In this episode of the Dynamics Corner Podcast, Kris and Brad speak with Søren Friis Alexandersen to discuss the Payables Agent in Microsoft Dynamics 365 Business Central and its role in automating invoice processing. They emphasize the importance of human oversight in AI decision-making and address the challenges of building trust in AI systems. The conversation underscores the importance of clear documentation to ensure effective policy implementation. By leveraging historical data and balancing coding with natural language, they outline how AI can enhance compliance and accounting practices. Additionally, they explore the importance of experimentation and contextual awareness to improve the effectiveness of AI agents. Ultimately, they argue that the future lies in the thoughtful integration of AI while ensuring that humans remain involved for accountability and transparency.
#MSDyn365BC #BusinessCentral #BC #DynamicsCorner
Follow Kris and Brad for more content:
https://matalino.io/bio
https://bprendergast.bio.link/
Mark as Played
Transcript
Episode Transcript
Available transcripts are automatically generated. Complete accuracy is not guaranteed.
Speaker 1 (00:00):
Welcome to another episode of Dynamics Corner
Agentic, trying to figure thatout what it means to you.
I'm your co-host, chris.
Speaker 2 (00:10):
And this is Brad.
This episode was recorded onApril 23rd 2025.
Chris, Chris, Chris, I liked tohear you play the, what you
call the ukulele.
I call it ukulele.
Ukulele.
It doesn't matter, it doesn'tmatter, it doesn't matter.
I've always wanted to hear youplay.
I know you have it, you play it, so we have to figure out.
One of these times I want youto maybe get a secondary mic or
(00:31):
something so you can play it andhear it too.
But you mentioned agents.
There's a lot of agents in thisworld.
There's travel agents, there'ssales agents, there's even AI
agents.
I'm not certain if you've heardof them, but today we did have
the opportunity to talk a lotabout the payables agents within
Business Central and what goesbehind it, what its features and
(00:51):
intended would be With us.
Today we had the opportunity tospeak with.
Good afternoon sir.
(01:12):
How are you doing?
Hello, nice to see you again.
Speaker 3 (01:16):
Good afternoon, good morning.
Good morning Is sound comingthrough, okay.
Speaker 2 (01:21):
The sound is coming through amazingly well.
That sounds amazing.
And I'm always fascinated byyour backgrounds is coming
through amazingly well.
That sounds amazing, and I'malways fascinated by your
backgrounds with the guitars aswell.
Well, you gotta play with uh,play on every string you got
right Like yeah you know, withso much going on, I always, I
often wish that I played aninstrument, but I don't have, or
(01:45):
I don't want to say that Idon't have the time, because
time is what you choose to dowith it, in my opinion.
Speaker 1 (01:51):
So it's you can still prioritize other things over.
Speaker 2 (01:55):
I want to play the piano or something.
Speaker 1 (01:57):
I think piano.
Speaker 2 (01:58):
I really enjoy piano music.
When my daughter was growing up, she played piano and we had a
piano in the house and that isprobably the one thing that
could occur that didn't annoy me.
Oh, one of the one things thatoccurred that couldn't annoy me.
Annoy me because she could playat any time and just hearing
the piano music is soothing.
Speaker 1 (02:18):
And I contemplated just See, right, you got to do
like a string instrument, thenyou could just learn four chords
and then you can probably playquite a bit.
Speaker 2 (02:29):
I think I'm scarred because when I was a young boy I
tried to learn to play theviola and I am left handed and
back when I was young theydidn't have a lot of left handed
devices for anything, forsports, for instruments.
The instruction didn't go toowell because they were trying to
(02:49):
teach me to play right-handedwhen I was holding it
left-handed, so the viola wasupside down in a sense, so it
just didn't work out too well.
Speaker 3 (03:01):
If I could switch one instrument.
I've often thought if I couldchoose any other instrument and
be equally good at thatinstrument just by clicking a
switch and trade off the guitaror something else, I'd go for
the saxophone.
I don't know Something in thesaxophone, that's just amazing.
Play some blues.
Yeah, it's always amazing.
(03:25):
I'm not complaining, I'm not.
Speaker 2 (03:30):
No, it's this.
I'm finding more.
I don't know if you get to acertain point.
There's so much available to doin life and I want to do it all
and it's very difficult tochoose.
Where do I focus my time sothat I can get the enjoyment
enjoyment even though I'd liketo do everything I need to hang
my stuff because I have themright.
Speaker 1 (03:52):
Oh wow, I got my ukuleles.
I got a bunch of them right onthe floor and I should be
getting them off the floor well,maybe we could have you do a
jam out session, chris, beforewe we finish.
Speaker 2 (04:05):
Uh yeah, the conversation.
How are things going in denmark?
The weather must be breakinggarden getting ready yes, uh,
amazing garden days.
Speaker 3 (04:16):
Today, a few days ago had a bit of rain.
So april is always.
April is always, uh, a bit ofeverything.
I mean you can ever everythingfrom almost summer like, and
then you can have everythingfrom almost summer-like, and
then you can have almostwinter-like some days.
In a few weeks it'll be morestable, but today, which happens
to be my birthday, happybirthday.
Speaker 2 (04:40):
Happy birthday.
It will be a little bit delayedwhen we play this, but happy
birthday.
Speaker 1 (04:43):
I hope that you have a great day to get to celebrate
and enjoy.
Speaker 2 (04:49):
Yes, I was thinking of you.
The other day I went to what Icall the farm store and I'm
toying with chickens again.
No, a real farm like a farmsupply store and we're a little
bit late for where I am rightnow for chicks.
But I was looking to see aboutgetting chicks again because I
had chickens for a long periodof time.
But I was thinking of you withthe farming and how chickens
(05:11):
would be great in your backyardthere.
Speaker 3 (05:14):
I would love to and we definitely want to Maybe not
this year, but I'm thinking nextyear we'll do it and see how it
goes.
Speaker 2 (05:24):
Oh, that's excellent.
That's one.
That's probably one of thethings that I can point back in
my life that I say I enjoyed themost was when I had the chicken
, so I think you'll enjoy it.
It's, um, and I even raisedsome uh the chickens uh
fertilize the rooster fertilized, uh, some of the eggs and one
of the chickens decided to siton them, so I was able to raise
chicks, a a few chicks, whichwas interesting to watch, but I
(05:47):
don't think we got together totalk about farming.
We'll save that for anotherconversation.
But before we jump into theconversation, would you mind
telling us a little bit aboutyourself?
Speaker 3 (05:56):
Yeah, so great to speak to you guys again.
So my name is Soren, I'm aproduct manager in the Business
Central engineering team and theresource development team, if
you will, based out ofCopenhagen, denmark, or just
north of Copenhagen, a smalltown called Lingby, and I've
been with Microsoft for nineyears and some months now,
(06:20):
almost all of them minus one anda half years in the Business
Central team, almost all of themminus one and a half years in
the business team Before that inMicrosoft Denmark in the sales
organization as an accounttechnology strategist.
So I have sort of a broad ITbackground.
I've been in IT the last 28years, so quite a few ER, you
(06:43):
know P systems under my beltstarted with Concord, xal, c5,
then the whole Navision Damguardthing happened and, yeah,
through, my first meeting withNav was with Nav 5, or that was
(07:03):
no 4, sorry my introduction.
Then I had a few touch pointswith some early versions, but
then everything from Nav 5 andupwards I've been working with
end customers.
I was self-employed, had a smallpartner business before joining
Microsoft where I was sort of80% engineer, 20% everything
(07:24):
else.
As you are, you'reself-employed.
Oh yeah, yes, speaking of timeto do everything, mostly doing
back then when I wasself-employed, mostly doing web
services integration from NAV toand from something else on the
other side, joined Microsoftpart of the sales organization
(07:45):
one and a half years and then,yeah, eight years now-ish in the
BC team, first doinglocalizations, rolling out BC to
many countries and thenswitching to some onboarding
five years ago now and did thatuntil almost a year ago and then
since about a year ago,everything now says AI, so
(08:09):
taking over the finance area ofBC since last fall and working
on the payables agent.
Speaker 2 (08:20):
The payables agent.
That opens up a wholeconversation for me.
With the payables agent we wereall at directions.
We had the opportunity to speak, we saw the keynote
presentations on the sales orderagent and there were other
discussions on the agents andyou also had a session on the
(08:41):
payables agent.
If you would, before I jumpinto it, I remember some of
those names.
Over at Microsoft, do they havea shrine to all the old
versions of Navision where yougo back from Navision 1?
Like an install kit orsomething on the wall where you
have all of the installed media?
Speaker 3 (09:00):
I don't know.
We have the physical media likethe DVDs and stuff and discs,
but we probably have all theartifacts on some shared drive
somewhere.
But you could say that so manypeople in the BC team have been
here for so long that the shrineis sort of made up of real
(09:23):
people that have just been withthe team since back in the day,
right.
So we have people who have beenwith the team like plus 20
years, like who remember youknow stuff way back from the you
know terminal-based client,even stuff like that.
I mean, you'd be amazed, yeah.
Speaker 2 (09:41):
No, definitely it's amazing and as you're talking
through it, it just it's, it's.
It spans quite a bit of time,but it doesn't feel like a long
time.
It feels like a lot of this wasjust yesterday it's.
Speaker 3 (09:55):
I just had that discussion.
It does feel like yesterday andit does.
It's also amazing that you canlike imagine bc now running in
the cloud in this super complexinfrastructure and services,
global scale, hyperscale cloud,on a system that was, or at
least inherited from a systemthat was basically built like 30
(10:18):
, 40 years ago almost, but atleast from a functionality
perspective there's stillremnants from that time and it
just goes to show that somebusiness processes don't change
that much.
Like core finance is still sortof core finance to a large
degree.
Of course there's some stuffyou want to do and stuff you
(10:39):
want to improve in yourbusinesses, but to the core,
we've been moving productsaround like as businesses more,
or to the core, we've beenmoving products around like as
businesses more or less the sameway, uh, for the last I don't
know many decades.
Uh, it is, and I think thatwhen I think a gl.
Speaker 2 (10:55):
when I think a gl entry, sales and receivable
setup, purchase and payablesetup, inventory setup, posting
groups, like a lot of that stuffhas been there, to me it feels
like the beginning, but I can'twait to speak to you.
We'll walk down memory lane thenext time we have a
conversation.
The Pables agent, if you would.
(11:17):
The whole world is talkingabout agents every day.
I don't think I can go withoutshutting everything off.
I don't think in the world thatyou can go an hour without
hearing about something AI oragent related and the payables
agents you have been workingwith.
Would you mind telling us alittle bit about the payables
agent?
Speaker 3 (11:37):
Yeah, no sure.
So, speaking of things thathaven't changed in decades,
companies have received invoicesfrom vendors, like since you
know or like all the time, andthat process, like receiving and
(12:00):
processing vendor invoices, hasalways been sort of prone to
automation.
We've been, you know, companieshave been trying to automate
that for decades, all the wayback to the 60s with EDI started
and, you know, as internetconnectivity got better and
various ways to get data intosystems, out of systems,
improved integrations, accesspoints, what have you.
(12:22):
So there's been sort of a pushto automate that process for
many decades, but it's alwayshit a roadblock when it hit that
last thing where now someoneneeds to actually process this
invoice and recognize what it is.
It's been very mechanical andif you received an invoice that
(12:44):
was based on, you had mapped upsome of the information or the
invoice to your GL accounts orto your items or things of that
nature.
Of course you've had OCR,optical character recognition,
for quite some time, but stillthat helped you with getting the
data into a sort ofmachine-readable format.
But after that, what wasactually on the invoice?
(13:05):
Like having a system that couldrecognize.
What is this even all about?
What is this invoice about,without having to map anything
and set anything up.
That's been the hard thing.
That's really where thePayables agent brings something
new, which is we believe thatnot only can the agent get all
the data in from your invoicesand we'll be starting with
(13:27):
normal PDF invoices but theagent will be able to recognize,
from an accounting perspective,what's going on on the invoice.
Is this a subscription thatwe've prepaid and thereby you
need to defer the expense, or isthis some other expense that we
have?
Something that's in our policythat says we need to post a
(13:50):
certain way or so all thataccounting knowledge, if you
will, we imagine the ABLE willbe able to acquire and utilize.
So basically helping theaccounting professional with
doing that last mileregistration of the invoice, if
you will, and in time, alsopurchase order matching and sort
(14:14):
of, in a sort of next phase.
But that accounting knowledgepiece has been the missing thing
, the hard thing to do until now, until now with LLMs, basically
.
So we think we can make thatwork.
That's what we're trying to doright now as we speak.
So we're building the agentright now and focusing first on
(14:36):
getting the PDF sourced, so froman email account or from a
SharePoint or OneDrive folderwhere you might want to place
your invoices, or from aSharePoint or OneDrive folder
where you might want to placeyour invoices, or from an email
account.
Like a trip scenario.
You forward or you get an emailwith an invoice from a vendor,
or an employee forwards aninvoice.
We import that into what wecall e-documents in BC and then
(14:58):
the agent takes over and startslooking at the invoice and sees,
based on a handful of differentways of processing this invoice
, it tries to look at how did wepreviously handle this invoice?
What does my purchase historysay in terms of how to post it,
how to register it, basicallyhow to code up the lines for the
(15:21):
invoices?
Do we have a policy in place?
Maybe you have a Word document,pdf document, that describes
your accounting policy.
When we post rent for thewarehouse, how should that line
look like?
Is there a certain dimensionthat should be added?
(15:41):
If you buy software licenses,does that need an allocation
account to be split up against,you know, based on multiple deal
accounts or dimensions orthings of that nature.
All of these things we'retrying to look at in sort of
combination and figure out sortof what is the best predictor
(16:03):
for how should this invoice behandled.
So that's really the idea andhopefully that will work out so
smooth that you can be basicallyhands-off.
Of course there'll be somecorner cases that are that are
hard for the agent to sort ofsay.
You know, for example, let'ssay you've been doing it a
certain way, posting rent forthe warehouse a certain way for
(16:26):
the last 10 months, but nowsuddenly last month you did it
differently.
Now what's the right answer?
Is it what you did last monthor does the 10 preceding months
weigh more heavily?
And so probably that's ahuman-in-the-loop moment where
you want to say okay and youknow.
So probably that's a human inthe loop moment where you want
(16:47):
to say okay, the agent can't doa confident enough, uh, can't
see confidence enough decision,so probably we need to involve a
human.
So human in the loop is a sortof a core concept.
Um, and everything that isagents, it's built around human
intervention, if you it is.
Speaker 2 (17:04):
I think of this when I hear the agents and get more
familiar with agents.
It's not just.
It seems like it's theagentification I always stumble
when I say that Agentificationof the world, because everyone
talks about agents to dospecific tasks and listening to
your explanation, it almost youstart to equate to how people
(17:25):
work in a sense, because if youhave an agent that's responsible
for receiving an invoice,taking that invoice or a
document invoice of some sort,taking that invoice and then
creating a purchase invoicewithin Business Central and then
processing it according to somerules, you're in a scenario
where, well, now we have adifference.
(17:48):
What do we happen in adifference?
It could even be the same casewhen you have a person doing
that same or a human doing thatsame task that they may need to
go speak to their supervisor, orthey may need to speak to
someone else to ensure accuracyof it, or there may be limits
that you can put into place.
Where you say the human andloop, I'm glad you explained
(18:09):
that because that's also anotherterm that I've been hearing
quite frequently inconversations, I think more so
in the business centralcommunity versus outside.
But what the human and loopmeans is, in essence where
humans get involved.
If I'm'm correct, humans getinvolved into the operation to
validate, confirm or be involvedin a process or a transaction.
Speaker 3 (18:36):
Yes, and when it comes to human in the loop, it's
it's super important thatthere's a it's very clear to the
user or users or team orcompany when is an agent
expected to do something andwhen is a human expected to do
something.
So there needs to be a veryclear handoff, if you will,
(18:58):
between well, the agent is nowrunning with the ball, but gets
stuck, and now it needs toinvolve a human.
So we've designed with the ballbut gets stuck, and now it
needs to involve a human.
So we've designed, as you'vemaybe seen, we've designed this
agent sidecar on the right side,where there's this timeline,
but the agent will describeanything that it's done in the
process and the end-to-end flow,basically, and then the human
(19:18):
will be brought in and will havea task to do if something is
uncertain, if the agent cannotcontinue, or if you've opted in
to be in the loop in a certainplace.
So human in the loop willdefinitely be like there will be
configurable human in the loopmoments where you can decide oh,
I always want to be in the loopwhen the vendor is not a good
(19:40):
match for what we have in oursystem, or the bank account
doesn't match, or whatever, orthere's an anomaly with the
amount.
It's suddenly twice as high asit used to be, things of that
nature.
But you want to maybe configurethat yourself, or maybe you
want to in the beginning to havelow autonomy or be brought into
the loop more until you feelcomfortable with the agent
(20:00):
taking some of those decisionsby itself, and then you dial up
the autonomy as you sort of goforward.
But there will also be otherhuman-in-the-loop moments where
we will decide well, there willbe human-in-the-loop when X
happens, for example.
That might be if it's sort ofmore risk situations or where
(20:21):
the agent simply cannot justmake a decision.
And how will the agent knowwhen to make a good decision and
I think that's a concept thatwe're working on is what is that
confidence level?
So, as I mentioned before, let'ssay you had the purchase
history, you had the policy youhad.
What if the agent suggestssomething?
It suggests a GL account topost this rent for the warehouse
(20:42):
, but for some reason the userdoesn't agree and changes that
to something else.
That signal that the userchanged it.
It's also a signal for theagent to say, aha, my suggestion
wasn't good enough.
I can remember this scenarionext time for this vendor for
this specific type of expense.
(21:02):
Correlate that with what doesthe policy say?
What does the policy say?
What does the history say?
Are you also using recurringpurchase lines?
Maybe that's also somethingthat should be taken into
account of how to weigh allthese things together.
So how we weigh these things.
That's then maybe the trickypart.
Should policy have priorityover history, or the other way
(21:23):
around?
And what about just pure LLMguesswork?
So if you don't have any ofthese things, you don't have any
history, you don't have anypolicy.
Just let the LLM guess.
Here's my chart of accounts.
Which one is the best forposting rent to my warehouse?
Probably the agent will havesome kind of idea about that,
because it does have some basicdomain knowledge, if you will.
Speaker 1 (21:45):
I was just going to ask you about that, about the
confidence level.
You know clearly the humanaspect of it, where you're
interacting with it and it'sinteracting with your business.
When do you have the authority?
Or maybe changing thatconfidence is like okay, well,
I'm not.
You know, it tells me that it's99% confident or 90% confident.
(22:08):
Maybe I'm not, I'm notconfident enough for that to
just make a decision.
Do you have control over that?
I think that becomes a uh,maybe, a need, uh from a
business aspect, like uh, maybe,maybe, when it maybe, when it's
only 90%, I don't want it to doanything, I want it to ask me,
(22:28):
right?
Speaker 3 (22:29):
That's an excellent question, Chris, and I think I
totally agree, and that controlfirst starts with transparency.
So what we're looking at firstis to so normally, a purchase
invoice.
So right now, what we do withthese invoices we take them in
and we create purchase invoice.
So right now, what we do withthese invoices we take them in
and we create purchase invoicedocuments from the MNPC, and
(22:49):
later we'll do PO matching andso on.
But and right now a purchaseorder sort of a purchase invoice
document is.
You could say it's already akind of a draft because it's not
posted yet, but we needed astage before that, a sort of a
draft stage before that, beforeit becomes a purchase invoice,
because someone might stumbleupon it and wouldn't know if
this invoice is ready forwhatever the next step is.
(23:11):
So we introduced a new purchasedocument draft stage, if you
will, where we create theinvoice based on all the
information we have, as wetalked about before, all the
input and signals to how do wethink this invoice should look
like as a draft and there theuser would look at it, and here
(23:32):
we plan to incorporate some UIthat says oh, we suggested this
GL account because you have thisin your policy.
Or we suggested this yieldaccount because that's the most
frequently used for this type ofexpense on this vendor.
So similar to what you see with.
You've seen the new auto fillfeature where it explains, with
(23:54):
small icons beside fields, whywas this predicted.
Imagine the same thinghappening here for all of the
information on invoice.
You could say, oh, it suggestedthis because it came from our
policy document and it's to somedegree certain about that.
Of course, when it comes toactual percentages, that's going
(24:16):
to be the tricky part.
We need to figure out how do welike if something comes from a
policy, are we then 95% certainor are we 85% certain?
That's something we'll have toexperiment with and build out
some models for certainty orconfidence.
That is not an easy feat, sothat's something we'll need to
(24:37):
figure out.
But we'll need to start with atleast explaining to the user
why these suggestions.
Why does this invoice look likethis on as detailed as level as
possible, and then you can takea decision on great, that's
fine.
You might also be able to sayin your configuration of the
agent that the policy shouldhave preference over history,
(25:02):
for example, because you mighthave done a certain thing for
posting rent the last 12 months,but now it's new fiscal year,
Now you change your policy forsome reason.
So now it shouldn't rely onhistory, it shouldn't be the
predominant signal.
So we imagine you can sort ofconfigure the hierarchy of how
much should they weigh thesefactors.
(25:25):
So it's not yeah, it'ssomething we need to get some
experimentation done with andget some feedback on.
We have this private previewprogram where we want to let you
guys and other people you knowlose on the agent and try it out
and see how does it work,Because you also have some
obviously you have a lot ofexperience with from from from
(25:45):
real customer scenarios howthese things work.
So, uh, so we're super excitedand we think we can make this
work.
Uh, but it's all about trust.
You need to be able to trustthe agent.
Definitely, you want to be ableto trust it before you dial up
the autonomy, and you also wantto be able to trust it before
you dial up the autonomy, andyou also want to be able to
trust it to just use it in thefirst place, like to get the
(26:09):
value from it, and I thinkthat's going to be the.
So, knowing accountants, ofcourse, you could say well,
you've been automating thisspace for years, but those
systems have been more sort offixed and based on mapping, so
you always knew what was goingto happen.
This is different, and our mainchallenge is to whenever AI is
(26:35):
in place, is to say, however weposted rent for the warehouse
yesterday, we need to suggestthe same way tomorrow.
It needs to be some consistency.
We can't live with the normalco-pilot creativity, if you will
.
That won't do in accountingright.
(26:56):
So that's what we're trying todo.
We're trying to tame the wildtiger of LLMs here in Incense
and get to some consistentresults, and that's probably the
hardest part of all of this.
Speaker 1 (27:12):
For sure, and I think there's a requirement for
businesses to really documentquite a bit, because we talk
about agents and grounding themto.
You know, with with somepolicies, as you had mentioned
and I chuckle internally alittle bit because any time that
I've done implementations and Iknow in the past there a lot of
(27:36):
businesses don't have actuallywritten policies, right, it's,
it's all like tribal knowledge.
And so you know, as we gotowards the agentic world, I
think it's more important nowfor organizations to really
write down and create thosedocumentations or create those
(27:56):
policies, so that when thesethings come out, you want to be
able to ground them based uponthose documentation, because
there's so many tribal knowledgeand you know we've all gone
through implementations or, likeGod, that person knows about
everything.
It's like how come you don'tknow?
Isn't it written?
No, they just know.
Right, it's tribal knowledge.
(28:16):
So I think it's a lot moreimportant now for businesses to
really take the time now forbusinesses to really take the
time.
Speaker 2 (28:27):
It is important.
It lessens the risk, becausethe knowledge that you're
talking about could be standardknowledge and the policies could
be governmental restrictionpolicies that you may have on
top of organizational policies.
So it's not like you have todraft everything, but you need
to have those sources that youcan reference, which are handy.
And, chris, I completely agreewith your point on so many times
(28:49):
a lot of policies, a lot ofprocedures and a lot of why
things are done are based upon.
Well, chris has been sittingthere for 20 years and that's
the way that he's always donethat and Chris may know the
reason.
But if somebody else doesn'tknow it, then when chris leaves
(29:12):
for whatever reason, theneveryone's stuck figuring out.
Either we're going to still doit this way we have no reason
why or we may do it a differentway.
I I'm listening to you discussthe agents and trust.
Trust becomes a big thing, butI I try to take it back of
agents will have a specific taskwhich I I want to talk about
where we come up with agents andhow we draw the line between
agents and use multiple agentswith handoffs, and you have to
(29:35):
have trust in them, but thatsame trust just because the
agents seem to be doing more, asyou said, not in like a fixed
automation but in a variableautomation because they can do a
lot of ad hoc reasoning.
I guess you could sayEveryone's always had to have
trust in systems because, lookat, if somebody's going to a new
ERP system from a spreadsheet,they have to go through a phase
(29:59):
of.
I need to trust that.
Whatever ERP system that we'reusing, you know from just my
passion, we'll say thateverybody's implementing
Business Central.
There is a sense of trust thatthey need to have there.
So I think it's also themessaging of a lot of this that
we do.
We already do with getting thattrust with, as I had mentioned,
(30:19):
you have the human in the loopwhere supervisors may come in
and do some checking.
And also, chris, to your pointof grounding as the agent
continues to do more work, ithas more information to ground
upon, just as if you were tobring on a new talented
individual to help you with yourpayables or purchases for the
(30:42):
agent.
And the key point to this isthe agents are tools to help
humans do their job, whateverthat job may be, to make it so
that they can spend their timeelsewhere, maybe in something
that's more valuable for them todo, more thought-provoking,
where an agent can't do the task, versus some of the repetitive
(31:04):
tasks of new purchase invoicevendor number 10,000 or 1000,
you know, purchased Athens chair, you know, for one it's.
It becomes like that.
So.
Speaker 3 (31:15):
No, that's, that's all true.
And actually, when it comes tothat trust so I'll come back and
just speak to the policy partuh, because there are some
things I didn't mention but butbefore I go there, uh, about
this trust, you're so right.
We, we've already today, placedthis trust in the humans, who
who do these tasks, and when II've spoken to a lot of cfos
(31:38):
lately and many of them say we,we actually have a hard time
finding qualified accountants,finding that qualified talent,
like just as humans.
So a lot of mistakes are beingmade in these departments today,
but maybe we have higherdemands, higher expectations for
(32:02):
agents and for AI than we havefor humans, or at least we're
not as forgiving when machinesmake mistakes as we are with
humans, because the machine hasto do it right and I expect it
to work, and otherwise it's abuggy thing, and if it doesn't
(32:22):
work, I'll maybe just abandon it, and will it ever a buggy thing
?
And if it doesn't work, I'llmaybe just abandon it, and will
it ever get a second chance?
Like?
We seem to be different towardsmachines, and that's something
we need to get over.
It's something we need to.
It's something that challengesour mindsets when working with
machines, but that is a new erathat we're in.
(32:42):
I think we all need to learnhow to work with these machines
in a way where they're also notconsistent always so and with
humans and not either, right.
So you could ask Peter theaccountant hey, how do I post
this rent to the GL?
I could ask him today, andchances are if I ask him in 30
(33:05):
days from now I might get adifferent answer.
But we don't think of it thatway.
But that's the reality.
He might have forgotten, orhe's maybe suddenly in doubt
because he slept in.
There's so many reasons whyhumans have deficiencies as well
.
Speaker 1 (33:23):
Sorry, sorry, it's just interesting.
You know we're talking abouttrust with the machines I can't
remember what movie it was where.
And there's a reason whyrobotics they put like human
faces, because then you have abetter connection, because
that's how we're designed right.
We're designed to like trustfaces, but when it comes to
machine one mistake I will nevertrust're designed right, we're
designed to like trust faces,but when it comes to machine,
(33:44):
one mistake I will never trustit again.
Right, like versus a humanperson.
I get it, you're human, you'llmake mistakes, but oh, you're a
machine.
You made a mistake.
I don't trust you because youcould make a mistake every
single time now, so that's-.
Speaker 2 (33:58):
No, exactly, that's fascinating.
I laugh because I say the samething and even with some of the
mistakes, I think some machinesare more consistent than humans.
In some tasks it's even yeah,it's a long way.
It is an interesting.
I think we'll get there.
I think we're in the infancyand I think a lot of it is the
(34:18):
fear, as far as fear, of whatare these agents going to do to
me, to my value, to what I need.
So I think some of it may bedriven from survival of I'm
against the machine.
That's not to come up with therage against the machine.
It's like the first thing Ithink of, but it's like you said
it's.
We don't have that level oftrust yet.
(34:39):
We have the trust in humans and, chrisris, I did see that as
well as is.
Speaker 3 (34:43):
That's why they put faces on it, because it's
supposed to be human, like withyou, which um yeah, no, that's
true and it's into, it'sinteresting, like we're it's and
this is a primal thing, likewe've had it.
It's just the way humans aresort of encoded like we need
faces, like that's also why, uh,our brain waves seem to sort of
(35:05):
sync up where we, when we're inthe same room, like when we
look each other in the eye,there's so many things that we
just that we don't see, that wedon't like understand, but our
brains just work in a very sortof, uh primal way in the sense,
like we're not, like we've we'veonly been on this planet for
such a short while, like interms of evolution, that the
(35:26):
things that, yeah, we're nottrying to do, that that
challenges, uh, challenge how we, how we work as humans.
Um, I want to go back to thepolicy thing because, uh, I just
want to say absolutely agreeabout companies not having their
policies written down, at leastnot the size of companies we
would typically work with.
(35:47):
Maybe some of the very largeones have some very well
documented and then again, inmany cases they don't.
It's still tribal knowledge.
We could also play with the ideathat if you have history, once
you enable the agent, you couldhave a function that would
(36:08):
generate a policy document basedon your history, right?
So use LLMs to generate a newpolicy document and say, hey,
this is like we're basicallydrafting up a work document that
says, based on yourpost-to-purchase history, this
seems to be your policy.
Go now and audit that and then,when it's ready, put that in
(36:30):
place.
And now you, of course, need tokeep that up to date and you
decide when you sync that backinto BC or whatever you do, or
let that be the new knowledgebase.
So we could play with thatthought as well.
Also, brad, you had aninteresting point about policy.
So it's not only like thecompany's own policy, it's also
regulation, governance, ofthings of that nature.
(36:53):
We also envision that in time,the agent will be able to say,
oh, I don't know how it is inthe US, but, for example, if you
buy IT equipment over a certainamount in Denmark, you need to
depreciate it over five years orthings of that nature.
So it needs to be a fixed asset.
So things of that nature likevery local to your business,
maybe to your industry, like ifit figures out that on the IRS
(37:17):
website it says something aboutyour industry and whatever an
expense.
However, a certain expenseneeds to be handled.
That could play a part as well.
That's further down the road,but we do want it to be your
accountant guide.
Basically, it should have thatknowledge in some way or form.
So those are some of the ideasfor the longer term.
Speaker 2 (37:42):
I think that's a great idea because it can be
where it's helpful, because thatcan help a business.
It goes with.
It's almost when we went fromhaving encyclopedias on the
shelf to having search engineswhere you could search for
information to be able to see itquicker.
But also now, if you can havean agent we'll talk about, since
we're talking about purchasesand payables or the payables
(38:04):
agents tongue-tied this morningit can help you make the right
decision so you don't risk anyviolations, because it also can
help you reduce your risk,because now you have these
policies and now you have anagent that can help you validate
that the accountant that youmay have, even if it's a new
regulation or a new rule.
(38:27):
I couldn't even keep up withwhat the IRS does over here.
I think it's purposeful howconfusing they make it.
But if you have some new rulesthat come in that, as you had
mentioned, that Peter theaccountant may not be fully
versed in, if he could havereference to an agent that can
bring it to light, to whereeither it reviews it or alerts
Peter to say, hey, there's a newrule or hey, this rule's here,
(38:48):
and then he could investigateand determine does this fit the
new rules or regulations?
Speaker 1 (38:57):
So it can be extremely beneficial in right
for agent Because there's beenan understanding or at least you
know how people perceive thispayables agent as just an
automation thing, and I thinkmaybe they're missing the point.
Where the value comes is to beable to adjust and reason hey,
there's an update for this,there's an update for your
(39:19):
policy.
We should update that policyBecause, again, another
component of that is a policychanges.
Nobody updates the document.
You know.
Maybe they update three monthslater because now there's an
updated policy, but the lastthree months they've been using
the old policy.
So I think that's where thevalue comes in for these agents
is to be able to look at otherareas and say, make some
suggestions and you know, andagain, like you said, brandon,
(39:44):
it's to be able to make betterdecisions without you having to
really interact or mayberesearch on your own.
It's going to know that, hey, Ishould check this area first to
make sure it's updated.
Speaker 3 (39:58):
Yeah, and there's there's.
So, brad, you mentioned likerules, like just as a phrase
just before, and I think this issomething that this is where
we've been really exploring thelast sort of six months is how
much of this should be based onrules that you set up somehow in
BC.
Like think, like your rules andoutlook, when x is equals one,
(40:23):
then do this.
When you know expense equalsrent, then do this.
But it just seemed likesomething new that you had to
maintain and set up and createand we really wanted to avoid
that if we could.
So that's's why we're nowleaning on your history.
Maybe recurring purchase lines,policy and policy I think I hope
(40:47):
will become a big part, becausepolicy can play a part in so
many other areas, like autofill.
Like, now you create a newvendor.
Well, what if it figures outwhile you're typing that you're
creating a new shipping providerand in your policy you have
something that says, oh, whenyou create shipping providers,
you need to fill out this andthis and this, or you need to
(41:10):
always use a certain postinggroup, or what, what have you
like?
There could be so manydifferent uses for policies
around the system for uh timeregistration, for almost
anything like the policy concept.
If it works, which we hope itwill, then there are so many
different applications that canuse that concept when you create
(41:33):
data, when you create documentsand so on, not just in purchase
, but in everywhere, basically.
Speaker 2 (41:43):
You have, I think of this and it sounds all logical
and what I I think of and Icould understand it and it makes
sense that I always get stuckon.
You know, maybe it's becauseunderstanding how llms work, how
do you come up with the idea?
Not how do you come up with theidea, not how do you come up
(42:05):
there, you come up with the idea.
Now you come up with the rulesand the logic, as we talked
about, to be able to look at thehistory, be able to look at
external documents and policies.
Put it all together.
So you have, you know, ifsomeone's going through to
design and develop an agent,when to determine that, an agent
(42:29):
that has, again, the variableunderstanding, the lack of
better terms versus the setupwhere you said you have a fixed
setup, turn on, turn off, dothis, do that to where you can
have that large language model,variability and, in essence,
deep thinking.
And then how do you implementit.
And then to also to my pointwith this, I think of all the
(42:53):
stuff that you're talking about.
So now you have an idea thatyou're working with.
Should this be an agent?
Should this not be?
Now I'm going to determine howto execute it, but to me it's
almost difficult to plan becausethis whole agent space in large
language models is changing sodrastically each week, so that
(43:14):
something that you may beplanning and working on today,
next week you may have to have acompletely different plan.
I'm not certain if I'marticulating this one, it's just
my mind is thinking throughthis whole process of putting
this all together while the busis still moving.
Speaker 3 (43:30):
In essence, that is exactly right and that is what
we're doing.
Technology is shifting so fastthese days that it's both good
and bad.
So what we thought would be alimitation right now may not be
a limitation tomorrow or nextweek even, which is great.
(43:53):
But there's a lot ofexperimentation and I think that
because this field is also sonew, like this field of large
language models and theavailability of it, like this
field of large language marvelsand the availability of it is
like hasn't like two and a halfyears ago?
It wasn't there, like none ofus knew anything about this two
and a half years ago.
So it's like it's so new.
We're all learning and we needto do a lot of experimentation.
(44:16):
That's also why we're trying toimplement sort of a ways to get
functionality out to peoplelike yourself, so you can test
on bits and pieces that are veryrough, so we can get some
feedback that are not bound tothe six months release cycle
that we have now, because that'ssimply too long time to wait
(44:38):
for pivoting and so on, so that,along with just doing a lot of
experimentation, pivoting and soon, so that along with just
doing a lot of experimentation,but there's, I mean, if I could
give an advice to people outthere who want to start building
agents or even just some AIlike Copilot features.
Just describing what you wantto do like write it down in a
(45:01):
document, just write it out.
Describing what you want to dolike write it down in a document
, just write it out For thepayables agent.
I wrote a two-page documentthat described what happens Like
oh, an email comes in with thePDF invoice, what does the agent
do and how do I envision itwill do it?
So once I had sort of a graspon that, then if I couldn't
describe that, then I didn'tknow my scenario well enough.
(45:23):
That scenario I wanted toautomate.
I just assumed at that pointthat it could be done.
I didn't know.
The next thing to do could be totest if Copilot or the LLM or
whatever technology that you'reusing, has some kind of base
understanding of the domainyou're building into.
So, in my case, has some kindof base understanding of the
(45:46):
domain you're building into.
So in my case, I wanted to seeif LLMs had sort of a base
understanding of accounting witha purchase invoice.
So I took a PDF of a randominvoice, sent it to Copilot and
said hey, I'm an accountant, Iwork in Business Central.
What on this invoice should Ibe mindful of when I register it
(46:06):
in my system?
And one of the lines saidsomething like support and
service for six months.
So that was my test, to see ifit knew about deferrals, for
example, and it did.
In the answer it said oh, Ilooked at this PDF invoice.
Here's the totals and here'sthe customer, blah, blah, blah,
blah.
Oh, it seems like there issomething you need to create
(46:26):
some kind of deferral forbecause it's a service and
support for six months.
That was my test of.
Does it have some kind of baseknowledge of this domain at all?
Because there might be domainsthat are so specific that it
wouldn't have enough trainingdata to know.
But this is such a well-known Imean, accounting on this level
is such a well-known space thatthe LLMs know quite a bit about
(46:52):
it.
So at least we have a goodstarting point for probably
understanding something likejust take a look at this invoice
, take a look at this policy,take a look at this policy,
remember this policy, rememberthis chart of accounts.
What's the right thing to dohere?
So that's sort of as a fallback.
(47:15):
So the way we think about this.
Let's start with the mostspecific stuff.
If you map your yield accountsto some text on the invoice,
maybe that overrides everything,because that's an explicit
mapping Great.
If you don't have that, maybewe'll look at the policy.
If you don't have a policy, wemight look at your history.
If you don't have history, wemight just default to.
(47:35):
What does the LLM just thinkjust off the bat, or that's my
thinking right now, but let'ssee how this plays out right?
So no for sure, and it's supertricky to build this.
To be honest.
I mean we're.
I mean this is hard, this ishard to build.
Speaker 2 (47:55):
I can't put my head around it.
I mean I can, but I can'tbecause there's a lot of moving
parts with access to data,access to external information,
as you talked about, with localpolicies, even knowing where you
are.
Because now being able to, it'sall mind-blowing to me, because
, differentiating between USpurchase rules in the state of
(48:17):
Washington versus the Denmark,you know purchase rules that you
may have.
I'm not even certain if you doit by country, by county or
region, or how you separate themin there, but that has to be
taken to ticket.
Speaker 1 (48:33):
You could specify that from a prompt perspective.
For example, if you build anagent, you kind of set the rules
of this is your baseline prompt?
You are an accounts payable,that is your role.
You are based in Washingtonstate.
Consider the policy inWashington state.
So that's kind of your ground,where it doesn't go anything
(48:54):
beyond that.
You know, don't worry about theother states, you are in
Washington state.
You know, adhere to thepolicies and all that stuff.
So once you get that ground andand they could still go crazy
too, I don't know LLMs, they canstill hallucinate Right, and
then they could like ah,washington, washington DC, is
that what you mean?
Speaker 3 (49:13):
Not Washington state, no, that's exactly right, and I
think that's also why if wewere just relying on the LLM by
itself, maybe that wouldn't cutit.
So but because we also havehistory and you might have your
own policy document, maybe thosethings in combination can help
qualify the suggestions that itcomes with.
(49:34):
But that is the thinking, likeexactly what you're saying there
, that in time it could havesort of a local view of the
world as well.
But it's also, but there's, oh,but there's so many challenges
that we meet there where we meetsome of the inherent things
that we're missing from atechnical perspective in BC.
So, for example, businessCentral is not inherently aware
(49:58):
of itself in a sense of what isa product in Business Central.
Business Central doesn't knowthat.
And a developer needs to tellBusiness Central search the item
table, and these are the seventables, right, that are related
to the item table, or how manyit is.
But if you have an extensionthat extends the item table, now
(50:19):
it's not just eight tables, nowit's nine or 10 tables that
make up a product in BC.
But Business Central doesn'tknow that.
I can say to BC search productsand then give me X, y, z.
So Business Central is notself-aware in that sense, that's
something we need to work on,to figure out how do we do that,
(50:40):
because we also need toconsider add-ons.
You might have extensions thatextend even your purchase lines
when you register an invoice.
You might have other fields onthe purchase lines that we don't
know from a Microsoftperspective, but the customers
added.
But we need to respect those aswell when we suggest the
purchase line for this newlyreceived invoice.
(51:00):
So there are many tricky things, those as well.
When we suggest the purchaseline for this newly received
invoice, right.
So there are many, many trickythings.
Then comes we also don't knowthe industry of this company.
Okay, that might be easier tosolve.
We could just add a new fieldto company information.
You fill it in, you know, basedon some predefined list, maybe,
(51:23):
or maybe we just look up yourcompany name online little and
figure out what is the industryof your company, something like
that, if that's reliable enough.
Then comes things that are sortof user specific or role
specific.
We don't.
So obviously we know what therole the payable agents has, but
we don't really know whoeverturned on the payables agent,
(51:44):
what their role is.
That might be a person wearingall hats in the company that
turned it on.
It might be the accountant whoturned it on or some admin, ic,
admin person.
So there's so many things wherewe need to get a lot more
context, and those are thingsthat we're we're trying to work
on to figure out how can we makeBC more self-aware and
(52:05):
context-aware.
But this is just under the hood.
So many things.
I want to come back to one thingyou mentioned, brad, so I think
well, both of you mentioned it.
Actually, no doubt we see anagent as having a specific role
or specific business process tosort of fulfill, like handle
(52:30):
this venture, invoice processfrom start to finish.
Already, there we're able todefine some context for this
agent, to define some bounds forit to say, I mean, if it tries
to do something, we can sort ofrein it in or fence it in in a
way, because we know the context.
We don't imagine one agent, onesuper agent, that just does
(52:53):
everything, has all permissionsin the system, that can do
everything.
An agent needs to have someinherent knowledge about what is
it there for, and so we're,yeah, and that also has, of
course, some benefits in termsof permissions, because it's
it's, it's a user in the byitself.
You can set up the, the uh,what is what it has access to in
(53:16):
pc, and so on, um, but in terms.
So come back to your questionbrad about how to stitch all
these things together.
What is AL code and where doesthe agent instructions fit in?
And when do we ask the LLM andwhere do we feed it data If it
needs to search our purchasehistory?
(53:37):
Can we just loop through 1million deal entries or like all
these things?
And this is super hard.
The one thing that we'reexperimenting with now, because
we have our own agent frameworkinside BC.
That is not Copilot Studio.
We went our own way because wewere waiting for Copilot Studio
(54:00):
to bring some of thecapabilities that we need.
So right now, the sales orderagent and payables agent in BC
are built inside BC, if you will.
Those agents work in a waywhere they can navigate the UI.
So a bit like the client, likethe business central client that
(54:23):
you see in UI, imagine theagent being able to navigate
that UI.
So let's take a scenario whereyou receive an invoice.
You need to figure out just asuper simple example.
You need to figure out if thisbank account that is stated on
the invoice is known to us onthe vendor.
(54:44):
We could do it in AL.
We could just do a three linesof code in AL check loop through
the vendor bank accounts andcheck if it's a match.
Yeah, but that doesn't soundvery agentic.
In a way we could also tell theagent go to the vendor card, go
(55:05):
to the vendor bank accounts, gothrough the bank accounts one
by one and compare them to thisbank account you have on the PDF
.
If we can make that work asinstructions, like in a no-code
instructions, hey, go check ifthe vendor bank account is known
to us by going to the vendorcard, blah, blah.
Go check if the vendor bankaccount is known to us by going
to the vendor card, blah, blah,blah.
(55:25):
Then we've not written code Now.
We've written natural languagein an instructions prompt
somewhere.
So we always have thesedecisions to make and the agent
say do we just code it in AL ordo we try to push the agent to
do it via the UI?
And we try to do an agenticfirst motion here.
Because if we can succeed withthat then in time you can
(55:48):
imagine just going into BCpulling up a window, describe
the agent that you want and howit should behave in the UI and
then it would just go do allthese things, no AL code needed.
Basically, it would justnavigate the UI.
That is sort of the thinking ofour agents within BC.
Speaker 2 (56:07):
You said a key phrase in there, a key few-choice
words, where you said with theAL code, we're writing code for
a specific scenario and then nowwe can instruct the agent to
work with the UI.
Now we can instruct the agentto work with the UI.
So is the agent working withthe UI directly to perform its
(56:29):
actions, or is it working withcode behind the scenes?
I guess you could say in alow-code fashion of concrete
scenarios that it's goingthrough and trying to find, and
then also to the to tack ontothat, what about the variability
?
I'm all for agents havingspecific tasks and then they can
(56:52):
do those tasks.
Well, similar to what we'retalking we talked about at the
onset, where there's so manythings that we want to do or I'd
like to do, but I can't do itall because I don't have time.
Nor could I do be you know, forlack of better terms good at
anything if I do too much,because I'll have no time to be
able to focus on it.
So agents can have specifictasks.
How is it working?
(57:14):
And what about variablescenarios?
So today we talked about bankaccount.
Okay, now we can put it in, wecan teach it, and I do want to
go back to the UI and how itinteracts with it.
But what if tomorrow it's ourvendor number for a vendor?
Speaker 3 (57:31):
Yeah, let's maybe go to the UI piece first, because
just a quick bit of explanationof how it works and I might I'll
probably vastly oversimplifythings here and if some of our
engineers are listening to thisthey'll probably say, yeah, it's
a bit more detailed than that.
So think of a page object withall the metadata it has about
(57:58):
which fields are on it and so on.
Today, very simply put, allthat data is being exposed to
the client as some internals orweb service or APIs.
The client reads that andrenders the page right.
So that same API the agent nowhas access to.
(58:21):
So the agent can just notrender but read all that same
information that's on the page.
So when we say navigate to thevendor's page, it does so like
the client would.
In essence, there's no physicalclient spun up somewhere, it's
just the agent consuming thosesame APIs and navigating.
Seeing in code it's like thematrix.
(58:43):
You see all that green text.
The agent basically just looksat all the metadata and say, oh,
there's a caption that saysvendors there, let me try to
navigate to that.
So that's very simple.
But the way the agent can sortof navigate the UI, you know,
can say, hey, go to the linesand create a new line, add this
value here, and that's also whyit will always see what the user
(59:07):
would see, if it has the access, of course, to that.
So that also means that when webuild UI for the agent to
traverse, we need to removeambiguity and need to remove
stuff that makes it confusingfor the agent, which happens to
have the nice side effect thatwhat is confusing for agents is
(59:30):
probably also confusing forusers.
So if we can simplify some more, that's probably a good thing.
Yeah, so that's just about howthe agent will sort of work.
Will we sometimes have to dothings in AL instead?
Have to do things in AL instead, where, because it's just like,
(59:50):
for example, finding a certainrecord or things of that nature,
might make more sense in somecases in AL instead of having
the agent do it?
That's also a cost concern orconsideration we need to think
about.
Well, when the agent navigatesthe UI in this way, it uses the
(01:00:12):
LLM, so inherently we will havea cost of running the agent,
basically because that wewouldn't, in the same way, have
an AL, at least not.
Well, the machines where theNST is running is running anyway
, but still there are manyfactors in this, in deciding do
(01:00:32):
we go for AL, do we go fornatural language?
In essence, and we're trying topush more to go towards natural
language.
Sorry if I didn't really answeryour question there, but no you
did answer it and it's.
Speaker 2 (01:00:43):
I'm pleased to hear there's also a consideration
today, because no one knows whatthe future would be.
It's almost like I remember whenthe cell phone was invented and
it was probably $300 a minuteto use the cell phones.
Now cell phones are fairlyreasonably priced for unlimited
text and data for a given monthin the onset, while the
(01:01:14):
technology is still new and itcan be costly.
In essence, to see that thereare some considerations to
appropriately use the agent inconsideration for cost and
function, not just to use anagent for the sake of using an
agent.
So it is nice to hear thatthose considerations are in
there and I do understand nowthat it's working with the ui
and that's how we can limit whatit has access to and what it
(01:01:34):
can do within the context of theuser executing the agent and
what they have access to, if I'mif I understood correctly yes,
it's I've had conversation withsome and there may be a
misconception that you knowagents are just, you know,
co-pilot workflows or powerautomate flows or some fashion
(01:01:59):
of that, and it's nice to hearthe details, and I've been
having conversations witheveryone.
It's a lot more than that and isthat a true misconception?
Or is it just somebody tryingto knock AI similar to back?
When I see Chrissy, as we saythis almost every day, it's you
use the appropriate tool forwhat you're trying to do.
Asking AI to say five plus fiveand it and being able to make
(01:02:21):
it do 11 doesn't do anythingbecause you can use a calculator
, or maybe there should be anagent for math versus an agent
for payables.
So it's speak.
It's just like people or humans.
I speak to a specific human ifI need specific information.
You know an electrician, acarpenter or something.
I went on a little tangent there, but yes, you did answer my
(01:02:41):
question.
Speaker 3 (01:02:43):
And I think that this is worth just drilling a bit
more into like the role of anagent.
I know we always spoke aboutthat it has a certain role, uh,
to, to, to play like.
Uh, how that?
Like, if you draw up anorganization and you look at the
agent landscape like, what,what should an agent landscape
(01:03:04):
look like in in thisorganization?
Oh, do I need a payables agent?
Do I need a sales agent?
Do I need, like, when is myplate full of agents that now
all bases are covered in mycompany?
And how do they hand over toeach other?
And do I have a master agent ofall of them?
Or do I have, like that sort oforchestrates it all or are they
(01:03:25):
able to orchestrate betweenthem and just hand over?
And how small should they be?
Like, also, when you considerthat it's a user that they sort
of emulate in a sense, should weconsider that the classic sort
of segregation of duties, allthose things like do an agent
map to a human one-to-one orlike, and how does that look and
(01:03:49):
how does that look like in thisorganization?
So, and there are so manythings like, we've only just
begun, so I think that'ssomething we need to get wise on
as we go along.
Personally, I don't see onesuper agent that does everything
.
I would rather go smalleragents, much more limited scope,
because that also means we candefine their context much more
(01:04:09):
precisely.
Much more limited scope,because that also means we can
define their context much moreprecisely.
So they're.
They ideally wouldn't go out ofbounds of what they're supposed
to do so often.
But then you could imaginehaving like, let's say, you had
five finance agents.
Then you could have anoverarching supervisor agent
within finance that could helpsort of orchestrate between
these five finance agents.
(01:04:30):
Could also be you had a vendorcommunication agent that said oh
, now we get an invoice in, thisgoes to the payables agent.
But the next time there's anemail from the vendor it happens
to be a reminder that theyhaven't gotten our payment.
So obviously that doesn't go tothe payables agent, this goes
to someone else, another agent.
(01:04:51):
So who's orchestrating that ifthey're all monitoring that same
inbox?
That's stuff we need to figureout.
But the discussions havestarted.
How do we?
What does the agent landscapelook like?
How do they hand over?
Where do the humans fit in Allof that?
Where do we have the oneoverview to rule them all?
Speaker 1 (01:05:13):
and all that stuff.
You need that business uh,orchestrated or that
orchestrates all the agents.
Right, like if, if I'm abusiness owner and I need it to
do something, I I I'd ratherjust go to a single agent that
can help me orchestrate toaccomplish the goal of my
(01:05:35):
business, rather than, yeah, mygoal involves this and then I
got to go talk to individualagents, versus going to a single
maybe a tenant agent that livesin your tenant and it says I
need to do this so it may pullin certain points of my
(01:05:58):
applications, whether that'sbusiness central or maybe I need
to interact with teams andwhatever that is, it's much
easier for me from a businessperspective, much easier to go
to a single agent that willorchestrate that.
Speaker 3 (01:06:09):
And that's a great way of thinking about it.
And I think when I say that weI personally I'd like to
componentize agents as much aspossible, try them as small as
possible.
That doesn't mean you'll have100 interaction points as a user
.
You could still have that oneinteraction point, as you say,
like the master orchestratoragent, or maybe within that
domain you work as a user, andthen that agent will, will,
(01:06:30):
coordinate with all the otherusers.
So it will, it will havedirects, if you will like.
Uh, so it's.
It's a bit the same kind oforganization, yeah, same
hierarchy, if you will.
And so, and I think that that's, that's just some
technicalities.
We need to figure out how doyou interact with an agent.
It doesn't necessarily need tobe a one-to-one representation
(01:06:52):
of how are the agentsrepresented in the system under
the hood somehow.
So, uh, you could have a one,one to twenty relationship, uh,
between you and agents, butthere could be a single pane of
glass that will be thatinterface or whatever that would
be super fascinating.
Speaker 1 (01:07:11):
It would almost really like it's a one-to-one
human, like a true co-pilot,right.
Like I'll have a co-pilot of aVP of operations, right.
Like that would be the personthat I would go to and say I
need to accomplish this, how doI accomplish this in the system.
And then that agent for thatspecific role for operations
(01:07:32):
would then just interact withall the different agents.
That also kind of follows thehierarchy within your
organization.
That's how maybe I'd see it,because then every role has a
co-pilot.
Speaker 3 (01:07:47):
No, to me that makes perfect sense.
And I think where we are now isthat we need to start with we
need to also be mindful of wherewe are and not try to build
100% automation to begin with.
We need to learn and there'sbenefits in involving users more
(01:08:11):
often to begin with, becausethere's a lot of learning
moments when users are involvedWhenever, as I mentioned 30
minutes ago, whenever the userchanges some of our suggestions
or the agent's suggestions,that's a learning opportunity.
So there's sort of great valuein having human in the loop and
I think if we can combine itwith the risk of being a little
(01:08:34):
bit philosophical but if we canmake humans and agents work
together and complement eachother's weak sides, that's that
we would have a verywell-working agent.
Aiming for 100, I think.
Think, of course I mean wewouldn't want accounts payables
to be 100% if it's reliable andfast and well, maybe speed
(01:08:58):
doesn't matter as much, but atleast reliable.
But I think that will take timeuntil we get there and that's
fine.
If we can solve for the 75% ofthe normal day-to-day use cases,
that's also super cool.
And then humans can do the morecomplex handle, the more
(01:09:20):
complex invoices and over time,as we get better using this
technology, the agents can getbetter.
Then probably we can increasethat to I don't know 80, 90
percent, who knows if we getclose to 100.
Speaker 2 (01:09:37):
There's so much in there and I think it's, you know
, the future.
Who knows where the future willbe?
And again, I do think we're atthe point just in our society,
just to go philosophical for alittle bit because,
realistically, some of this AIagent stuff, with it being so
new and learning what it can doand seeing how it reacts,
responds and works, isphilosophical because we don't
(01:10:01):
really know.
But it would just beinteresting to see where it will
all be and how this will allwork uh, in the years.
And I think now we're justthrowing a lot of agents against
the wall to see which work,which stick, which are effective
, and then also how to, to yourpoint, maybe even break them
down to have separate agentsthat have specialties or
(01:10:25):
specializations, to be able tohave them work together.
Speaker 1 (01:10:29):
You know, I think it's going to be.
I would love to see it more of arole specific, because, if you
look at the way I use agents orco-pilot, right, for example, in
my day-to-day it's typicallyaround my role, and so it would
be interesting to see if this ismaybe where you guys are
(01:10:52):
heading toward, where Microsoftis going towards, where there is
a co-pilot working along withme, because then I can have my
conversation with my CEO or yourpresident or anybody else in
your organization is morestrategic, strategic, right.
There's something that us humansare more creative and more
(01:11:13):
strategic about things, that weyou know what we do in the
business, where the hard, youknow, the tedious work of
getting to do things is wherethe co-pilot comes in, where,
hey, I need some data, you know,give this information to me
while we're trying to make adecision of where the business
is going.
I think that would make a lotmore sense that you'd always
have a co-pilot with you, basedon the role and, of course,
(01:11:35):
structurally in terms ofhierarchy within the system.
It also makes sense because weunderstand that your agents are
based upon a role, right,because it's UI driven.
I think it's limited to aspecific role, and so maybe
that's where it is, because nowyou have, you know, an AP
payables agent, now you didaccounts receivable agent and so
(01:11:56):
forth, sales agent and so forth.
Maybe that is where thestructure is.
To me it makes more sensebecause my area of focus is just
that's, my interaction anyway.
Speaker 3 (01:12:09):
No, and I think that makes perfect sense, and I think
that this is why we've allchosen this path and chosen sort
of well-known scenarios, roleseven you can say they're even
more narrow than a role at thispoint, because a payables agent,
like a real accountant, mightdo more, or an AP clerk might do
(01:12:29):
more than just registerinvoices.
I guess it depends also percompany, and I think what we're
doing now is a lot ofexperimentation, and I think
that's the only way forward.
As you said, brad, like we tryto create some agency which ones
work as the technology improves, as it does every day, and I
think we also just need to behonest that I mean, as I said,
(01:12:52):
two and a half years ago thisfield didn't exist, so there are
, by definition, no experts.
We're also learning, right,we're just working on this all
day, every day, so we get a lotof exposure to it as
technologists.
So we get a lot of exposure toit as technologists, but experts
(01:13:13):
, I mean I would like to saywe're getting there, but you
know who can be an expert in twoand a half years of exposure to
something?
So this will take time, sowe're all learning right and an
expert on something that'schanging daily.
Speaker 2 (01:13:28):
So it's not only that it's new.
Generally speaking, it's in itsinfancy.
But it seems every time I wakeup and put on the morning
finance news, they're talkingabout some new LLM, local large
language model, some new agent,some new version that it's very
difficult for anyone to keep up,which is also important to
(01:13:51):
mention.
Some are talking about that itmay be too late to learn or too
late to get into it, buteverybody's in the same boat
because everything's new.
So no one has missed the agentexperimentation phase, even
Copilot Studio, as you hadmentioned.
Speaker 3 (01:14:10):
they keep adding features and changes, so and get
in and start working with ityeah, and this is, of course, a
much larger challenge that we,like the industry, face, and all
companies out there wants toconsume all companies out there
who wants to consume AI or toconsume features within their
(01:14:32):
ERP system or whatever.
So, on one side, we have ahigher than ever before desire
to experiment, test out things AB test, try out things with
real customers and so on, andthat will only increase Our
(01:14:53):
desire to do that will onlyincrease, and we need that
because technology is changingEvery day.
We look in the toolbox, toolslook different, work different.
So on the other side of that,you have customers who need
predictability, they don't wantto be disrupted, they value
business continuity, all thesethings.
(01:15:13):
So how do these two ends meet?
We need experimentation andwhat you call fail fast or adapt
fast or pivot fast Customers.
They just wanted to work thesame way as they did yesterday.
So how do they meet?
And I think that comes back to,at least for Business Central
(01:15:33):
specifically, we need to figureout ways where we can create
test rings and let people optinto being early preview testers
or maybe test experimentalfeatures in a sandbox or what
have you, so we can get somesome feedback.
Yeah, but this is, this is achallenge, for sure.
(01:15:54):
And also when it comes toadoption, as we spoke about an
hour ago, like if it doesn'twork or not in terms of a bug,
but if I don't get the resultfor of a feature that I expected
of this agent, will I abandonit or will I give it a new shot
tomorrow where we've maybechanged some stuff and updated
(01:16:16):
it and all of that.
You had a great episoderecently about change management
, which I loved.
That was just.
All of these things are justmore important than ever, and I
think we have some part of thatthat we need to do Educate,
explain value even in the UI,make it evident what is the
(01:16:40):
value for you as a person, notonly for the company, but also
for the company.
This is going to be even moreimportant than ever and super
hard.
Speaker 1 (01:16:51):
Yeah, I think I mean just thinking about just going
back to the hierarchy componentand you brought up the adoption.
You know, a lot of times youknow employees or users of an
application or even within thebusiness itself.
It typically bleeds from whatyou do on a personal side of
things and then you kind ofbring that into your business.
(01:17:11):
So, for example, I'd call that.
You know I use Copilot orChatGPT or whatnot and when you
chat right, you have differentchats.
Typically, when you're creatingyour chat you ground it of what
that role is.
You are my finance manager andI have this situation and I need
(01:17:36):
you.
It's almost like that, right,with Business Central.
I would think that's wherewe're heading.
I can't foresee that.
I'm just thinking out loud herewhere I would love to structure
it that way where what I do nowpersonally I could do the same
thing in Business Central.
I'm going to add an agent.
You are this agent.
I need you to do this.
(01:17:57):
It's almost in that way of easyfor people to kind of
understand because that's whatthey do.
Now I mean, it's becoming aday-to-day occurrence.
I chat with my chat GPT or grokor whatever AI that you use for
chat.
But again, I can't predict that.
Just the way I do things in aday-to-day, it may bleed into
(01:18:21):
Business Central.
I don't know.
Speaker 3 (01:18:24):
Yeah, but I think it makes sense.
And if you think about the waythat the agent, like you, have a
bunch of agents enabled foryour organization.
They work for different teams,for different users and so on,
and at the end of the day, therewould be supervisors of agents,
like human supervisors ofagents, that would work with
(01:18:47):
agents, make sure that thewheel's always spinning, that
the agents are not stuck onanything, of course, with all
the business approvals built inthat you typically want.
When you need to approve a newvendor that's great in the
system, or need to send the POfor approval, all that stuff
still works as it used to.
(01:19:07):
The agents don't change that.
All that stuff still works asit used to right the agents
don't change that.
It might be that you can haveagents that will help these
approvers as well.
Like you can imagine an agentthat says you know, here's a new
approval request from Chris whowants to approve this purchase
order of $10,000.
Well, based on history, there'snothing to worry about.
(01:19:30):
You can easily go approve thisone.
Should I go approve it the nexttime on your behalf?
There are many applicationswhere you can think about agents
kicking in and learning fromthe history and, looking at you
know, try to identify risk,because if there's no risk, why
stop at all?
Why not just go ahead and andum, yeah, but let's see how far
(01:19:53):
we can.
We can drive this automationbecause it is automation brad
you asked about.
Was this just a glorified flow?
Depends on how you look at it.
I mean, from a conceptualstandpoint, you could say yeah,
uh, but it's a flow that's ableto handle much more variability,
in a sense, like where flow isvery deterministic.
(01:20:15):
You put a stick in the wheel ona certain point it just breaks
and falls apart.
Where here it's a bit so yeah.
A guy said to me some months agowell, this is like when you
build houses in areas where youhave a lot of earthquakes you
(01:20:35):
build them so the foundation cansort of move and the house can
move.
It's a bit like the same thing.
You need to build in somewiggle room so the agent can
take some decisions within acertain context, within certain
constraints and so on, but notbreak when something doesn't
follow the sunshine pathUnderstood.
(01:20:57):
It's not a bunch of ifs andthens yes, it's.
Speaker 2 (01:21:02):
No, it's.
I mean, I guess, ultimately, Iguess any workflow is a bunch of
ifs and thens.
It's just a matter of which ifsand thens and whens you can
react to, or respond to withoutbreaking.
So we've talked a lot about thepayables agent.
What is the current state ofthe payables agent as far as
(01:21:24):
availability and what are theshort-term future plans?
I know nobody can plan too farinto the future because
something could change tomorrow,but to anybody listening, we've
been talking about it.
What is the current state andavailability for it?
Speaker 3 (01:21:39):
Yeah, great question.
So because the agent navigatesthe UI, as we've talked about,
we've had to improve some stuffin sort of the base application,
especially around e-documents.
We've had to build a connectorwhere basically JobQ, or the
agent, via the JobQ, could pullin an invoice that comes in on a
(01:22:03):
certain mailbox or on aSharePoint folder and so on.
So we've done some componentsthat the agent will use and
those are right now, and we'llcall sort of private preview
where if you're on a certainlist, your tenant is on a list,
then you can get access to someof these components and try them
out.
This is not in any way anyagentic experience yet.
(01:22:26):
This is just some improvementsto the base product and we'll
keep doing those over the nextsort of few months and the idea
is that we'll add the first sortof version of the agentic
experience, enabling theend-to-end flow, creating the
invoice and so on, with the nextmajor release.
That is the next.
(01:22:47):
That is the plan.
So end of year, this calendaryear, basically, how much will
it then support in terms of howmuch accounting knowledge we'll
have?
Well, that's what we need tofigure out from now and until
then.
Will we have the policy conceptworking?
(01:23:08):
Will we have some learningconcepts working Well?
How much would we look atpurchase history when we try to
determine what the invoiceshould look like, and so on?
So that's to be decided.
But ideally we would have thefirst version working with the
next major release and in US andmaybe some other
(01:23:31):
English-speaking countries tobegin with, and then it's going
to be an iterative process,adding more capabilities, adding
more countries and so on afterthat.
But yeah, keep an eye out forthe next major release.
Speaker 2 (01:23:48):
That's the current plan Always a lot of great
things added to each of themajor releases, and well, so we
appreciate you taking the timeto speak with us today.
It's always a pleasure speakingwith you, whether it's either
on the podcast or in person, oreven in some remote
conversations.
We do appreciate you takingyour time to speak with us and
we appreciate all that you dofor the community and as well as
(01:24:10):
all of what your team is doingfor Business Central for the
users of the application.
If anyone has any questions orwould like to learn more about
the Payables, agents or otherinitiatives that you're working
on, what's the best way to getin contact with you?
Speaker 3 (01:24:22):
Well, this moment there's probably two good ways.
So one is on the Yammer site orViva Engage If you're a partner
.
Reach out there If you havequestions that you don't want
the world to see.
Otherwise, feel free to hook upwith me on LinkedIn.
It's the only place that I'msort of active from a sort of
social media perspective, sofeel free to send me messages
(01:24:44):
there or tag me in the post orwhatever, and I'll be there.
Speaker 2 (01:24:46):
Great, Thank you very much for taking the time to
speak with us again, and I hopeyou get some great gardening
into the season as well.
Speaker 3 (01:24:55):
Thank you so much, guys, pleasure all right, thank
you so much soon ciao, ciao, bye.
Speaker 2 (01:25:00):
Thank you, chris, for your time for another episode
of in the dynamics, corner chair, and thank you to our guests
for participating thank you,brad, for your.
Speaker 1 (01:25:09):
It is a wonderful episode of Dynamics Corner Chair
.
I would also like to thank ourguests for joining us.
Thank you for all of ourlisteners tuning in as well.
You can find Brad atdeveloperlifecom, that is
D-V-L-P-R-L-I-F-E dot com, andyou can interact with them via
(01:25:29):
Twitter D-V-L-P-R-L-I-F-E dotcom, and you can interact with
them via TwitterD-V-L-P-R-L-I-F-E.
You can also find me atMattalinoio, m-a-t-a-l-i-n-o dot
I-O, and my Twitter handle isMattalino16.
And you can see those linksdown below in the show notes.
(01:25:50):
Again, thank you everyone.
Thank you and take care.
Welcome to another episode of Dynamics Corner
Agentic, trying to figure thatout what it means to you.
I'm your co-host, chris.
Speaker 2 (00:10):
And this is Brad.
This episode was recorded onApril 23rd 2025.
Chris, Chris, Chris, I liked tohear you play the, what you
call the ukulele.
I call it ukulele.
Ukulele.
It doesn't matter, it doesn'tmatter, it doesn't matter.
I've always wanted to hear youplay.
I know you have it, you play it, so we have to figure out.
One of these times I want youto maybe get a secondary mic or
(00:31):
something so you can play it andhear it too.
But you mentioned agents.
There's a lot of agents in thisworld.
There's travel agents, there'ssales agents, there's even AI
agents.
I'm not certain if you've heardof them, but today we did have
the opportunity to talk a lotabout the payables agents within
Business Central and what goesbehind it, what its features and
(00:51):
intended would be With us.
Today we had the opportunity tospeak with.
Good afternoon sir.
(01:12):
How are you doing?
Hello, nice to see you again.
Speaker 3 (01:16):
Good afternoon, good morning.
Good morning Is sound comingthrough, okay.
Speaker 2 (01:21):
The sound is coming through amazingly well.
That sounds amazing.
And I'm always fascinated byyour backgrounds is coming
through amazingly well.
That sounds amazing, and I'malways fascinated by your
backgrounds with the guitars aswell.
Well, you gotta play with uh,play on every string you got
right Like yeah you know, withso much going on, I always, I
often wish that I played aninstrument, but I don't have, or
(01:45):
I don't want to say that Idon't have the time, because
time is what you choose to dowith it, in my opinion.
Speaker 1 (01:51):
So it's you can still prioritize other things over.
Speaker 2 (01:55):
I want to play the piano or something.
Speaker 1 (01:57):
I think piano.
Speaker 2 (01:58):
I really enjoy piano music.
When my daughter was growing up, she played piano and we had a
piano in the house and that isprobably the one thing that
could occur that didn't annoy me.
Oh, one of the one things thatoccurred that couldn't annoy me.
Annoy me because she could playat any time and just hearing
the piano music is soothing.
Speaker 1 (02:18):
And I contemplated just See, right, you got to do
like a string instrument, thenyou could just learn four chords
and then you can probably playquite a bit.
Speaker 2 (02:29):
I think I'm scarred because when I was a young boy I
tried to learn to play theviola and I am left handed and
back when I was young theydidn't have a lot of left handed
devices for anything, forsports, for instruments.
The instruction didn't go toowell because they were trying to
(02:49):
teach me to play right-handedwhen I was holding it
left-handed, so the viola wasupside down in a sense, so it
just didn't work out too well.
Speaker 3 (03:01):
If I could switch one instrument.
I've often thought if I couldchoose any other instrument and
be equally good at thatinstrument just by clicking a
switch and trade off the guitaror something else, I'd go for
the saxophone.
I don't know Something in thesaxophone, that's just amazing.
Play some blues.
Yeah, it's always amazing.
(03:25):
I'm not complaining, I'm not.
Speaker 2 (03:30):
No, it's this.
I'm finding more.
I don't know if you get to acertain point.
There's so much available to doin life and I want to do it all
and it's very difficult tochoose.
Where do I focus my time sothat I can get the enjoyment
enjoyment even though I'd liketo do everything I need to hang
my stuff because I have themright.
Speaker 1 (03:52):
Oh wow, I got my ukuleles.
I got a bunch of them right onthe floor and I should be
getting them off the floor well,maybe we could have you do a
jam out session, chris, beforewe we finish.
Speaker 2 (04:05):
Uh yeah, the conversation.
How are things going in denmark?
The weather must be breakinggarden getting ready yes, uh,
amazing garden days.
Speaker 3 (04:16):
Today, a few days ago had a bit of rain.
So april is always.
April is always, uh, a bit ofeverything.
I mean you can ever everythingfrom almost summer like, and
then you can have everythingfrom almost summer-like, and
then you can have almostwinter-like some days.
In a few weeks it'll be morestable, but today, which happens
to be my birthday, happybirthday.
Speaker 2 (04:40):
Happy birthday.
It will be a little bit delayedwhen we play this, but happy
birthday.
Speaker 1 (04:43):
I hope that you have a great day to get to celebrate
and enjoy.
Speaker 2 (04:49):
Yes, I was thinking of you.
The other day I went to what Icall the farm store and I'm
toying with chickens again.
No, a real farm like a farmsupply store and we're a little
bit late for where I am rightnow for chicks.
But I was looking to see aboutgetting chicks again because I
had chickens for a long periodof time.
But I was thinking of you withthe farming and how chickens
(05:11):
would be great in your backyardthere.
Speaker 3 (05:14):
I would love to and we definitely want to Maybe not
this year, but I'm thinking nextyear we'll do it and see how it
goes.
Speaker 2 (05:24):
Oh, that's excellent.
That's one.
That's probably one of thethings that I can point back in
my life that I say I enjoyed themost was when I had the chicken
, so I think you'll enjoy it.
It's, um, and I even raisedsome uh the chickens uh
fertilize the rooster fertilized, uh, some of the eggs and one
of the chickens decided to siton them, so I was able to raise
chicks, a a few chicks, whichwas interesting to watch, but I
(05:47):
don't think we got together totalk about farming.
We'll save that for anotherconversation.
But before we jump into theconversation, would you mind
telling us a little bit aboutyourself?
Speaker 3 (05:56):
Yeah, so great to speak to you guys again.
So my name is Soren, I'm aproduct manager in the Business
Central engineering team and theresource development team, if
you will, based out ofCopenhagen, denmark, or just
north of Copenhagen, a smalltown called Lingby, and I've
been with Microsoft for nineyears and some months now,
(06:20):
almost all of them minus one anda half years in the Business
Central team, almost all of themminus one and a half years in
the business team Before that inMicrosoft Denmark in the sales
organization as an accounttechnology strategist.
So I have sort of a broad ITbackground.
I've been in IT the last 28years, so quite a few ER, you
(06:43):
know P systems under my beltstarted with Concord, xal, c5,
then the whole Navision Damguardthing happened and, yeah,
through, my first meeting withNav was with Nav 5, or that was
(07:03):
no 4, sorry my introduction.
Then I had a few touch pointswith some early versions, but
then everything from Nav 5 andupwards I've been working with
end customers.
I was self-employed, had a smallpartner business before joining
Microsoft where I was sort of80% engineer, 20% everything
(07:24):
else.
As you are, you'reself-employed.
Oh yeah, yes, speaking of timeto do everything, mostly doing
back then when I wasself-employed, mostly doing web
services integration from NAV toand from something else on the
other side, joined Microsoftpart of the sales organization
(07:45):
one and a half years and then,yeah, eight years now-ish in the
BC team, first doinglocalizations, rolling out BC to
many countries and thenswitching to some onboarding
five years ago now and did thatuntil almost a year ago and then
since about a year ago,everything now says AI, so
(08:09):
taking over the finance area ofBC since last fall and working
on the payables agent.
Speaker 2 (08:20):
The payables agent.
That opens up a wholeconversation for me.
With the payables agent we wereall at directions.
We had the opportunity to speak, we saw the keynote
presentations on the sales orderagent and there were other
discussions on the agents andyou also had a session on the
(08:41):
payables agent.
If you would, before I jumpinto it, I remember some of
those names.
Over at Microsoft, do they havea shrine to all the old
versions of Navision where yougo back from Navision 1?
Like an install kit orsomething on the wall where you
have all of the installed media?
Speaker 3 (09:00):
I don't know.
We have the physical media likethe DVDs and stuff and discs,
but we probably have all theartifacts on some shared drive
somewhere.
But you could say that so manypeople in the BC team have been
here for so long that the shrineis sort of made up of real
(09:23):
people that have just been withthe team since back in the day,
right.
So we have people who have beenwith the team like plus 20
years, like who remember youknow stuff way back from the you
know terminal-based client,even stuff like that.
I mean, you'd be amazed, yeah.
Speaker 2 (09:41):
No, definitely it's amazing and as you're talking
through it, it just it's, it's.
It spans quite a bit of time,but it doesn't feel like a long
time.
It feels like a lot of this wasjust yesterday it's.
Speaker 3 (09:55):
I just had that discussion.
It does feel like yesterday andit does.
It's also amazing that you canlike imagine bc now running in
the cloud in this super complexinfrastructure and services,
global scale, hyperscale cloud,on a system that was, or at
least inherited from a systemthat was basically built like 30
(10:18):
, 40 years ago almost, but atleast from a functionality
perspective there's stillremnants from that time and it
just goes to show that somebusiness processes don't change
that much.
Like core finance is still sortof core finance to a large
degree.
Of course there's some stuffyou want to do and stuff you
(10:39):
want to improve in yourbusinesses, but to the core,
we've been moving productsaround like as businesses more,
or to the core, we've beenmoving products around like as
businesses more or less the sameway, uh, for the last I don't
know many decades.
Uh, it is, and I think thatwhen I think a gl.
Speaker 2 (10:55):
when I think a gl entry, sales and receivable
setup, purchase and payablesetup, inventory setup, posting
groups, like a lot of that stuffhas been there, to me it feels
like the beginning, but I can'twait to speak to you.
We'll walk down memory lane thenext time we have a
conversation.
The Pables agent, if you would.
(11:17):
The whole world is talkingabout agents every day.
I don't think I can go withoutshutting everything off.
I don't think in the world thatyou can go an hour without
hearing about something AI oragent related and the payables
agents you have been workingwith.
Would you mind telling us alittle bit about the payables
agent?
Speaker 3 (11:37):
Yeah, no sure.
So, speaking of things thathaven't changed in decades,
companies have received invoicesfrom vendors, like since you
know or like all the time, andthat process, like receiving and
(12:00):
processing vendor invoices, hasalways been sort of prone to
automation.
We've been, you know, companieshave been trying to automate
that for decades, all the wayback to the 60s with EDI started
and, you know, as internetconnectivity got better and
various ways to get data intosystems, out of systems,
improved integrations, accesspoints, what have you.
(12:22):
So there's been sort of a pushto automate that process for
many decades, but it's alwayshit a roadblock when it hit that
last thing where now someoneneeds to actually process this
invoice and recognize what it is.
It's been very mechanical andif you received an invoice that
(12:44):
was based on, you had mapped upsome of the information or the
invoice to your GL accounts orto your items or things of that
nature.
Of course you've had OCR,optical character recognition,
for quite some time, but stillthat helped you with getting the
data into a sort ofmachine-readable format.
But after that, what wasactually on the invoice?
(13:05):
Like having a system that couldrecognize.
What is this even all about?
What is this invoice about,without having to map anything
and set anything up.
That's been the hard thing.
That's really where thePayables agent brings something
new, which is we believe thatnot only can the agent get all
the data in from your invoicesand we'll be starting with
(13:27):
normal PDF invoices but theagent will be able to recognize,
from an accounting perspective,what's going on on the invoice.
Is this a subscription thatwe've prepaid and thereby you
need to defer the expense, or isthis some other expense that we
have?
Something that's in our policythat says we need to post a
(13:50):
certain way or so all thataccounting knowledge, if you
will, we imagine the ABLE willbe able to acquire and utilize.
So basically helping theaccounting professional with
doing that last mileregistration of the invoice, if
you will, and in time, alsopurchase order matching and sort
(14:14):
of, in a sort of next phase.
But that accounting knowledgepiece has been the missing thing
, the hard thing to do until now, until now with LLMs, basically
.
So we think we can make thatwork.
That's what we're trying to doright now as we speak.
So we're building the agentright now and focusing first on
(14:36):
getting the PDF sourced, so froman email account or from a
SharePoint or OneDrive folderwhere you might want to place
your invoices, or from aSharePoint or OneDrive folder
where you might want to placeyour invoices, or from an email
account.
Like a trip scenario.
You forward or you get an emailwith an invoice from a vendor,
or an employee forwards aninvoice.
We import that into what wecall e-documents in BC and then
(14:58):
the agent takes over and startslooking at the invoice and sees,
based on a handful of differentways of processing this invoice
, it tries to look at how did wepreviously handle this invoice?
What does my purchase historysay in terms of how to post it,
how to register it, basicallyhow to code up the lines for the
(15:21):
invoices?
Do we have a policy in place?
Maybe you have a Word document,pdf document, that describes
your accounting policy.
When we post rent for thewarehouse, how should that line
look like?
Is there a certain dimensionthat should be added?
(15:41):
If you buy software licenses,does that need an allocation
account to be split up against,you know, based on multiple deal
accounts or dimensions orthings of that nature.
All of these things we'retrying to look at in sort of
combination and figure out sortof what is the best predictor
(16:03):
for how should this invoice behandled.
So that's really the idea andhopefully that will work out so
smooth that you can be basicallyhands-off.
Of course there'll be somecorner cases that are that are
hard for the agent to sort ofsay.
You know, for example, let'ssay you've been doing it a
certain way, posting rent forthe warehouse a certain way for
(16:26):
the last 10 months, but nowsuddenly last month you did it
differently.
Now what's the right answer?
Is it what you did last monthor does the 10 preceding months
weigh more heavily?
And so probably that's ahuman-in-the-loop moment where
you want to say okay and youknow.
So probably that's a human inthe loop moment where you want
(16:47):
to say okay, the agent can't doa confident enough, uh, can't
see confidence enough decision,so probably we need to involve a
human.
So human in the loop is a sortof a core concept.
Um, and everything that isagents, it's built around human
intervention, if you it is.
Speaker 2 (17:04):
I think of this when I hear the agents and get more
familiar with agents.
It's not just.
It seems like it's theagentification I always stumble
when I say that Agentificationof the world, because everyone
talks about agents to dospecific tasks and listening to
your explanation, it almost youstart to equate to how people
(17:25):
work in a sense, because if youhave an agent that's responsible
for receiving an invoice,taking that invoice or a
document invoice of some sort,taking that invoice and then
creating a purchase invoicewithin Business Central and then
processing it according to somerules, you're in a scenario
where, well, now we have adifference.
(17:48):
What do we happen in adifference?
It could even be the same casewhen you have a person doing
that same or a human doing thatsame task that they may need to
go speak to their supervisor, orthey may need to speak to
someone else to ensure accuracyof it, or there may be limits
that you can put into place.
Where you say the human andloop, I'm glad you explained
(18:09):
that because that's also anotherterm that I've been hearing
quite frequently inconversations, I think more so
in the business centralcommunity versus outside.
But what the human and loopmeans is, in essence where
humans get involved.
If I'm'm correct, humans getinvolved into the operation to
validate, confirm or be involvedin a process or a transaction.
Speaker 3 (18:36):
Yes, and when it comes to human in the loop, it's
it's super important thatthere's a it's very clear to the
user or users or team orcompany when is an agent
expected to do something andwhen is a human expected to do
something.
So there needs to be a veryclear handoff, if you will,
(18:58):
between well, the agent is nowrunning with the ball, but gets
stuck, and now it needs toinvolve a human.
So we've designed with the ballbut gets stuck, and now it
needs to involve a human.
So we've designed, as you'vemaybe seen, we've designed this
agent sidecar on the right side,where there's this timeline,
but the agent will describeanything that it's done in the
process and the end-to-end flow,basically, and then the human
(19:18):
will be brought in and will havea task to do if something is
uncertain, if the agent cannotcontinue, or if you've opted in
to be in the loop in a certainplace.
So human in the loop willdefinitely be like there will be
configurable human in the loopmoments where you can decide oh,
I always want to be in the loopwhen the vendor is not a good
(19:40):
match for what we have in oursystem, or the bank account
doesn't match, or whatever, orthere's an anomaly with the
amount.
It's suddenly twice as high asit used to be, things of that
nature.
But you want to maybe configurethat yourself, or maybe you
want to in the beginning to havelow autonomy or be brought into
the loop more until you feelcomfortable with the agent
(20:00):
taking some of those decisionsby itself, and then you dial up
the autonomy as you sort of goforward.
But there will also be otherhuman-in-the-loop moments where
we will decide well, there willbe human-in-the-loop when X
happens, for example.
That might be if it's sort ofmore risk situations or where
(20:21):
the agent simply cannot justmake a decision.
And how will the agent knowwhen to make a good decision and
I think that's a concept thatwe're working on is what is that
confidence level?
So, as I mentioned before, let'ssay you had the purchase
history, you had the policy youhad.
What if the agent suggestssomething?
It suggests a GL account topost this rent for the warehouse
(20:42):
, but for some reason the userdoesn't agree and changes that
to something else.
That signal that the userchanged it.
It's also a signal for theagent to say, aha, my suggestion
wasn't good enough.
I can remember this scenarionext time for this vendor for
this specific type of expense.
(21:02):
Correlate that with what doesthe policy say?
What does the policy say?
What does the history say?
Are you also using recurringpurchase lines?
Maybe that's also somethingthat should be taken into
account of how to weigh allthese things together.
So how we weigh these things.
That's then maybe the trickypart.
Should policy have priorityover history, or the other way
(21:23):
around?
And what about just pure LLMguesswork?
So if you don't have any ofthese things, you don't have any
history, you don't have anypolicy.
Just let the LLM guess.
Here's my chart of accounts.
Which one is the best forposting rent to my warehouse?
Probably the agent will havesome kind of idea about that,
because it does have some basicdomain knowledge, if you will.
Speaker 1 (21:45):
I was just going to ask you about that, about the
confidence level.
You know clearly the humanaspect of it, where you're
interacting with it and it'sinteracting with your business.
When do you have the authority?
Or maybe changing thatconfidence is like okay, well,
I'm not.
You know, it tells me that it's99% confident or 90% confident.
(22:08):
Maybe I'm not, I'm notconfident enough for that to
just make a decision.
Do you have control over that?
I think that becomes a uh,maybe, a need, uh from a
business aspect, like uh, maybe,maybe, when it maybe, when it's
only 90%, I don't want it to doanything, I want it to ask me,
(22:28):
right?
Speaker 3 (22:29):
That's an excellent question, Chris, and I think I
totally agree, and that controlfirst starts with transparency.
So what we're looking at firstis to so normally, a purchase
invoice.
So right now, what we do withthese invoices we take them in
and we create purchase invoice.
So right now, what we do withthese invoices we take them in
and we create purchase invoicedocuments from the MNPC, and
(22:49):
later we'll do PO matching andso on.
But and right now a purchaseorder sort of a purchase invoice
document is.
You could say it's already akind of a draft because it's not
posted yet, but we needed astage before that, a sort of a
draft stage before that, beforeit becomes a purchase invoice,
because someone might stumbleupon it and wouldn't know if
this invoice is ready forwhatever the next step is.
(23:11):
So we introduced a new purchasedocument draft stage, if you
will, where we create theinvoice based on all the
information we have, as wetalked about before, all the
input and signals to how do wethink this invoice should look
like as a draft and there theuser would look at it, and here
(23:32):
we plan to incorporate some UIthat says oh, we suggested this
GL account because you have thisin your policy.
Or we suggested this yieldaccount because that's the most
frequently used for this type ofexpense on this vendor.
So similar to what you see with.
You've seen the new auto fillfeature where it explains, with
(23:54):
small icons beside fields, whywas this predicted.
Imagine the same thinghappening here for all of the
information on invoice.
You could say, oh, it suggestedthis because it came from our
policy document and it's to somedegree certain about that.
Of course, when it comes toactual percentages, that's going
(24:16):
to be the tricky part.
We need to figure out how do welike if something comes from a
policy, are we then 95% certainor are we 85% certain?
That's something we'll have toexperiment with and build out
some models for certainty orconfidence.
That is not an easy feat, sothat's something we'll need to
(24:37):
figure out.
But we'll need to start with atleast explaining to the user
why these suggestions.
Why does this invoice look likethis on as detailed as level as
possible, and then you can takea decision on great, that's
fine.
You might also be able to sayin your configuration of the
agent that the policy shouldhave preference over history,
(25:02):
for example, because you mighthave done a certain thing for
posting rent the last 12 months,but now it's new fiscal year,
Now you change your policy forsome reason.
So now it shouldn't rely onhistory, it shouldn't be the
predominant signal.
So we imagine you can sort ofconfigure the hierarchy of how
much should they weigh thesefactors.
(25:25):
So it's not yeah, it'ssomething we need to get some
experimentation done with andget some feedback on.
We have this private previewprogram where we want to let you
guys and other people you knowlose on the agent and try it out
and see how does it work,Because you also have some
obviously you have a lot ofexperience with from from from
(25:45):
real customer scenarios howthese things work.
So, uh, so we're super excitedand we think we can make this
work.
Uh, but it's all about trust.
You need to be able to trustthe agent.
Definitely, you want to be ableto trust it before you dial up
the autonomy, and you also wantto be able to trust it before
you dial up the autonomy, andyou also want to be able to
trust it to just use it in thefirst place, like to get the
(26:09):
value from it, and I thinkthat's going to be the.
So, knowing accountants, ofcourse, you could say well,
you've been automating thisspace for years, but those
systems have been more sort offixed and based on mapping, so
you always knew what was goingto happen.
This is different, and our mainchallenge is to whenever AI is
(26:35):
in place, is to say, however weposted rent for the warehouse
yesterday, we need to suggestthe same way tomorrow.
It needs to be some consistency.
We can't live with the normalco-pilot creativity, if you will
.
That won't do in accountingright.
(26:56):
So that's what we're trying todo.
We're trying to tame the wildtiger of LLMs here in Incense
and get to some consistentresults, and that's probably the
hardest part of all of this.
Speaker 1 (27:12):
For sure, and I think there's a requirement for
businesses to really documentquite a bit, because we talk
about agents and grounding themto.
You know, with with somepolicies, as you had mentioned
and I chuckle internally alittle bit because any time that
I've done implementations and Iknow in the past there a lot of
(27:36):
businesses don't have actuallywritten policies, right, it's,
it's all like tribal knowledge.
And so you know, as we gotowards the agentic world, I
think it's more important nowfor organizations to really
write down and create thosedocumentations or create those
(27:56):
policies, so that when thesethings come out, you want to be
able to ground them based uponthose documentation, because
there's so many tribal knowledgeand you know we've all gone
through implementations or, likeGod, that person knows about
everything.
It's like how come you don'tknow?
Isn't it written?
No, they just know.
Right, it's tribal knowledge.
(28:16):
So I think it's a lot moreimportant now for businesses to
really take the time now forbusinesses to really take the
time.
Speaker 2 (28:27):
It is important.
It lessens the risk, becausethe knowledge that you're
talking about could be standardknowledge and the policies could
be governmental restrictionpolicies that you may have on
top of organizational policies.
So it's not like you have todraft everything, but you need
to have those sources that youcan reference, which are handy.
And, chris, I completely agreewith your point on so many times
(28:49):
a lot of policies, a lot ofprocedures and a lot of why
things are done are based upon.
Well, chris has been sittingthere for 20 years and that's
the way that he's always donethat and Chris may know the
reason.
But if somebody else doesn'tknow it, then when chris leaves
(29:12):
for whatever reason, theneveryone's stuck figuring out.
Either we're going to still doit this way we have no reason
why or we may do it a differentway.
I I'm listening to you discussthe agents and trust.
Trust becomes a big thing, butI I try to take it back of
agents will have a specific taskwhich I I want to talk about
where we come up with agents andhow we draw the line between
agents and use multiple agentswith handoffs, and you have to
(29:35):
have trust in them, but thatsame trust just because the
agents seem to be doing more, asyou said, not in like a fixed
automation but in a variableautomation because they can do a
lot of ad hoc reasoning.
I guess you could sayEveryone's always had to have
trust in systems because, lookat, if somebody's going to a new
ERP system from a spreadsheet,they have to go through a phase
(29:59):
of.
I need to trust that.
Whatever ERP system that we'reusing, you know from just my
passion, we'll say thateverybody's implementing
Business Central.
There is a sense of trust thatthey need to have there.
So I think it's also themessaging of a lot of this that
we do.
We already do with getting thattrust with, as I had mentioned,
(30:19):
you have the human in the loopwhere supervisors may come in
and do some checking.
And also, chris, to your pointof grounding as the agent
continues to do more work, ithas more information to ground
upon, just as if you were tobring on a new talented
individual to help you with yourpayables or purchases for the
(30:42):
agent.
And the key point to this isthe agents are tools to help
humans do their job, whateverthat job may be, to make it so
that they can spend their timeelsewhere, maybe in something
that's more valuable for them todo, more thought-provoking,
where an agent can't do the task, versus some of the repetitive
(31:04):
tasks of new purchase invoicevendor number 10,000 or 1000,
you know, purchased Athens chair, you know, for one it's.
It becomes like that.
So.
Speaker 3 (31:15):
No, that's, that's all true.
And actually, when it comes tothat trust so I'll come back and
just speak to the policy partuh, because there are some
things I didn't mention but butbefore I go there, uh, about
this trust, you're so right.
We, we've already today, placedthis trust in the humans, who
who do these tasks, and when II've spoken to a lot of cfos
(31:38):
lately and many of them say we,we actually have a hard time
finding qualified accountants,finding that qualified talent,
like just as humans.
So a lot of mistakes are beingmade in these departments today,
but maybe we have higherdemands, higher expectations for
(32:02):
agents and for AI than we havefor humans, or at least we're
not as forgiving when machinesmake mistakes as we are with
humans, because the machine hasto do it right and I expect it
to work, and otherwise it's abuggy thing, and if it doesn't
(32:22):
work, I'll maybe just abandon it, and will it ever a buggy thing
?
And if it doesn't work, I'llmaybe just abandon it, and will
it ever get a second chance?
Like?
We seem to be different towardsmachines, and that's something
we need to get over.
It's something we need to.
It's something that challengesour mindsets when working with
machines, but that is a new erathat we're in.
(32:42):
I think we all need to learnhow to work with these machines
in a way where they're also notconsistent always so and with
humans and not either, right.
So you could ask Peter theaccountant hey, how do I post
this rent to the GL?
I could ask him today, andchances are if I ask him in 30
(33:05):
days from now I might get adifferent answer.
But we don't think of it thatway.
But that's the reality.
He might have forgotten, orhe's maybe suddenly in doubt
because he slept in.
There's so many reasons whyhumans have deficiencies as well
.
Speaker 1 (33:23):
Sorry, sorry, it's just interesting.
You know we're talking abouttrust with the machines I can't
remember what movie it was where.
And there's a reason whyrobotics they put like human
faces, because then you have abetter connection, because
that's how we're designed right.
We're designed to like trustfaces, but when it comes to
machine one mistake I will nevertrust're designed right, we're
designed to like trust faces,but when it comes to machine,
(33:44):
one mistake I will never trustit again.
Right, like versus a humanperson.
I get it, you're human, you'llmake mistakes, but oh, you're a
machine.
You made a mistake.
I don't trust you because youcould make a mistake every
single time now, so that's-.
Speaker 2 (33:58):
No, exactly, that's fascinating.
I laugh because I say the samething and even with some of the
mistakes, I think some machinesare more consistent than humans.
In some tasks it's even yeah,it's a long way.
It is an interesting.
I think we'll get there.
I think we're in the infancyand I think a lot of it is the
(34:18):
fear, as far as fear, of whatare these agents going to do to
me, to my value, to what I need.
So I think some of it may bedriven from survival of I'm
against the machine.
That's not to come up with therage against the machine.
It's like the first thing Ithink of, but it's like you said
it's.
We don't have that level oftrust yet.
(34:39):
We have the trust in humans and, chrisris, I did see that as
well as is.
Speaker 3 (34:43):
That's why they put faces on it, because it's
supposed to be human, like withyou, which um yeah, no, that's
true and it's into, it'sinteresting, like we're it's and
this is a primal thing, likewe've had it.
It's just the way humans aresort of encoded like we need
faces, like that's also why, uh,our brain waves seem to sort of
(35:05):
sync up where we, when we're inthe same room, like when we
look each other in the eye,there's so many things that we
just that we don't see, that wedon't like understand, but our
brains just work in a very sortof, uh primal way in the sense,
like we're not, like we've we'veonly been on this planet for
such a short while, like interms of evolution, that the
(35:26):
things that, yeah, we're nottrying to do, that that
challenges, uh, challenge how we, how we work as humans.
Um, I want to go back to thepolicy thing because, uh, I just
want to say absolutely agreeabout companies not having their
policies written down, at leastnot the size of companies we
would typically work with.
(35:47):
Maybe some of the very largeones have some very well
documented and then again, inmany cases they don't.
It's still tribal knowledge.
We could also play with the ideathat if you have history, once
you enable the agent, you couldhave a function that would
(36:08):
generate a policy document basedon your history, right?
So use LLMs to generate a newpolicy document and say, hey,
this is like we're basicallydrafting up a work document that
says, based on yourpost-to-purchase history, this
seems to be your policy.
Go now and audit that and then,when it's ready, put that in
(36:30):
place.
And now you, of course, need tokeep that up to date and you
decide when you sync that backinto BC or whatever you do, or
let that be the new knowledgebase.
So we could play with thatthought as well.
Also, brad, you had aninteresting point about policy.
So it's not only like thecompany's own policy, it's also
regulation, governance, ofthings of that nature.
(36:53):
We also envision that in time,the agent will be able to say,
oh, I don't know how it is inthe US, but, for example, if you
buy IT equipment over a certainamount in Denmark, you need to
depreciate it over five years orthings of that nature.
So it needs to be a fixed asset.
So things of that nature likevery local to your business,
maybe to your industry, like ifit figures out that on the IRS
(37:17):
website it says something aboutyour industry and whatever an
expense.
However, a certain expenseneeds to be handled.
That could play a part as well.
That's further down the road,but we do want it to be your
accountant guide.
Basically, it should have thatknowledge in some way or form.
So those are some of the ideasfor the longer term.
Speaker 2 (37:42):
I think that's a great idea because it can be
where it's helpful, because thatcan help a business.
It goes with.
It's almost when we went fromhaving encyclopedias on the
shelf to having search engineswhere you could search for
information to be able to see itquicker.
But also now, if you can havean agent we'll talk about, since
we're talking about purchasesand payables or the payables
(38:04):
agents tongue-tied this morningit can help you make the right
decision so you don't risk anyviolations, because it also can
help you reduce your risk,because now you have these
policies and now you have anagent that can help you validate
that the accountant that youmay have, even if it's a new
regulation or a new rule.
(38:27):
I couldn't even keep up withwhat the IRS does over here.
I think it's purposeful howconfusing they make it.
But if you have some new rulesthat come in that, as you had
mentioned, that Peter theaccountant may not be fully
versed in, if he could havereference to an agent that can
bring it to light, to whereeither it reviews it or alerts
Peter to say, hey, there's a newrule or hey, this rule's here,
(38:48):
and then he could investigateand determine does this fit the
new rules or regulations?
Speaker 1 (38:57):
So it can be extremely beneficial in right
for agent Because there's beenan understanding or at least you
know how people perceive thispayables agent as just an
automation thing, and I thinkmaybe they're missing the point.
Where the value comes is to beable to adjust and reason hey,
there's an update for this,there's an update for your
(39:19):
policy.
We should update that policyBecause, again, another
component of that is a policychanges.
Nobody updates the document.
You know.
Maybe they update three monthslater because now there's an
updated policy, but the lastthree months they've been using
the old policy.
So I think that's where thevalue comes in for these agents
is to be able to look at otherareas and say, make some
suggestions and you know, andagain, like you said, brandon,
(39:44):
it's to be able to make betterdecisions without you having to
really interact or mayberesearch on your own.
It's going to know that, hey, Ishould check this area first to
make sure it's updated.
Speaker 3 (39:58):
Yeah, and there's there's.
So, brad, you mentioned likerules, like just as a phrase
just before, and I think this issomething that this is where
we've been really exploring thelast sort of six months is how
much of this should be based onrules that you set up somehow in
BC.
Like think, like your rules andoutlook, when x is equals one,
(40:23):
then do this.
When you know expense equalsrent, then do this.
But it just seemed likesomething new that you had to
maintain and set up and createand we really wanted to avoid
that if we could.
So that's's why we're nowleaning on your history.
Maybe recurring purchase lines,policy and policy I think I hope
(40:47):
will become a big part, becausepolicy can play a part in so
many other areas, like autofill.
Like, now you create a newvendor.
Well, what if it figures outwhile you're typing that you're
creating a new shipping providerand in your policy you have
something that says, oh, whenyou create shipping providers,
you need to fill out this andthis and this, or you need to
(41:10):
always use a certain postinggroup, or what, what have you
like?
There could be so manydifferent uses for policies
around the system for uh timeregistration, for almost
anything like the policy concept.
If it works, which we hope itwill, then there are so many
different applications that canuse that concept when you create
(41:33):
data, when you create documentsand so on, not just in purchase
, but in everywhere, basically.
Speaker 2 (41:43):
You have, I think of this and it sounds all logical
and what I I think of and Icould understand it and it makes
sense that I always get stuckon.
You know, maybe it's becauseunderstanding how llms work, how
do you come up with the idea?
Not how do you come up with theidea, not how do you come up
(42:05):
there, you come up with the idea.
Now you come up with the rulesand the logic, as we talked
about, to be able to look at thehistory, be able to look at
external documents and policies.
Put it all together.
So you have, you know, ifsomeone's going through to
design and develop an agent,when to determine that, an agent
(42:29):
that has, again, the variableunderstanding, the lack of
better terms versus the setupwhere you said you have a fixed
setup, turn on, turn off, dothis, do that to where you can
have that large language model,variability and, in essence,
deep thinking.
And then how do you implementit.
And then to also to my pointwith this, I think of all the
(42:53):
stuff that you're talking about.
So now you have an idea thatyou're working with.
Should this be an agent?
Should this not be?
Now I'm going to determine howto execute it, but to me it's
almost difficult to plan becausethis whole agent space in large
language models is changing sodrastically each week, so that
(43:14):
something that you may beplanning and working on today,
next week you may have to have acompletely different plan.
I'm not certain if I'marticulating this one, it's just
my mind is thinking throughthis whole process of putting
this all together while the busis still moving.
Speaker 3 (43:30):
In essence, that is exactly right and that is what
we're doing.
Technology is shifting so fastthese days that it's both good
and bad.
So what we thought would be alimitation right now may not be
a limitation tomorrow or nextweek even, which is great.
(43:53):
But there's a lot ofexperimentation and I think that
because this field is also sonew, like this field of large
language models and theavailability of it, like this
field of large language marvelsand the availability of it is
like hasn't like two and a halfyears ago?
It wasn't there, like none ofus knew anything about this two
and a half years ago.
So it's like it's so new.
We're all learning and we needto do a lot of experimentation.
(44:16):
That's also why we're trying toimplement sort of a ways to get
functionality out to peoplelike yourself, so you can test
on bits and pieces that are veryrough, so we can get some
feedback that are not bound tothe six months release cycle
that we have now, because that'ssimply too long time to wait
(44:38):
for pivoting and so on, so that,along with just doing a lot of
experimentation, pivoting and soon, so that along with just
doing a lot of experimentation,but there's, I mean, if I could
give an advice to people outthere who want to start building
agents or even just some AIlike Copilot features.
Just describing what you wantto do like write it down in a
(45:01):
document, just write it out.
Describing what you want to dolike write it down in a document
, just write it out For thepayables agent.
I wrote a two-page documentthat described what happens Like
oh, an email comes in with thePDF invoice, what does the agent
do and how do I envision itwill do it?
So once I had sort of a graspon that, then if I couldn't
describe that, then I didn'tknow my scenario well enough.
(45:23):
That scenario I wanted toautomate.
I just assumed at that pointthat it could be done.
I didn't know.
The next thing to do could be totest if Copilot or the LLM or
whatever technology that you'reusing, has some kind of base
understanding of the domainyou're building into.
So, in my case, has some kindof base understanding of the
(45:46):
domain you're building into.
So in my case, I wanted to seeif LLMs had sort of a base
understanding of accounting witha purchase invoice.
So I took a PDF of a randominvoice, sent it to Copilot and
said hey, I'm an accountant, Iwork in Business Central.
What on this invoice should Ibe mindful of when I register it
(46:06):
in my system?
And one of the lines saidsomething like support and
service for six months.
So that was my test, to see ifit knew about deferrals, for
example, and it did.
In the answer it said oh, Ilooked at this PDF invoice.
Here's the totals and here'sthe customer, blah, blah, blah,
blah.
Oh, it seems like there issomething you need to create
(46:26):
some kind of deferral forbecause it's a service and
support for six months.
That was my test of.
Does it have some kind of baseknowledge of this domain at all?
Because there might be domainsthat are so specific that it
wouldn't have enough trainingdata to know.
But this is such a well-known Imean, accounting on this level
is such a well-known space thatthe LLMs know quite a bit about
(46:52):
it.
So at least we have a goodstarting point for probably
understanding something likejust take a look at this invoice
, take a look at this policy,take a look at this policy,
remember this policy, rememberthis chart of accounts.
What's the right thing to dohere?
So that's sort of as a fallback.
(47:15):
So the way we think about this.
Let's start with the mostspecific stuff.
If you map your yield accountsto some text on the invoice,
maybe that overrides everything,because that's an explicit
mapping Great.
If you don't have that, maybewe'll look at the policy.
If you don't have a policy, wemight look at your history.
If you don't have history, wemight just default to.
(47:35):
What does the LLM just thinkjust off the bat, or that's my
thinking right now, but let'ssee how this plays out right?
So no for sure, and it's supertricky to build this.
To be honest.
I mean we're.
I mean this is hard, this ishard to build.
Speaker 2 (47:55):
I can't put my head around it.
I mean I can, but I can'tbecause there's a lot of moving
parts with access to data,access to external information,
as you talked about, with localpolicies, even knowing where you
are.
Because now being able to, it'sall mind-blowing to me, because
, differentiating between USpurchase rules in the state of
(48:17):
Washington versus the Denmark,you know purchase rules that you
may have.
I'm not even certain if you doit by country, by county or
region, or how you separate themin there, but that has to be
taken to ticket.
Speaker 1 (48:33):
You could specify that from a prompt perspective.
For example, if you build anagent, you kind of set the rules
of this is your baseline prompt?
You are an accounts payable,that is your role.
You are based in Washingtonstate.
Consider the policy inWashington state.
So that's kind of your ground,where it doesn't go anything
(48:54):
beyond that.
You know, don't worry about theother states, you are in
Washington state.
You know, adhere to thepolicies and all that stuff.
So once you get that ground andand they could still go crazy
too, I don't know LLMs, they canstill hallucinate Right, and
then they could like ah,washington, washington DC, is
that what you mean?
Speaker 3 (49:13):
Not Washington state, no, that's exactly right, and I
think that's also why if wewere just relying on the LLM by
itself, maybe that wouldn't cutit.
So but because we also havehistory and you might have your
own policy document, maybe thosethings in combination can help
qualify the suggestions that itcomes with.
(49:34):
But that is the thinking, likeexactly what you're saying there
, that in time it could havesort of a local view of the
world as well.
But it's also, but there's, oh,but there's so many challenges
that we meet there where we meetsome of the inherent things
that we're missing from atechnical perspective in BC.
So, for example, businessCentral is not inherently aware
(49:58):
of itself in a sense of what isa product in Business Central.
Business Central doesn't knowthat.
And a developer needs to tellBusiness Central search the item
table, and these are the seventables, right, that are related
to the item table, or how manyit is.
But if you have an extensionthat extends the item table, now
(50:19):
it's not just eight tables, nowit's nine or 10 tables that
make up a product in BC.
But Business Central doesn'tknow that.
I can say to BC search productsand then give me X, y, z.
So Business Central is notself-aware in that sense, that's
something we need to work on,to figure out how do we do that,
(50:40):
because we also need toconsider add-ons.
You might have extensions thatextend even your purchase lines
when you register an invoice.
You might have other fields onthe purchase lines that we don't
know from a Microsoftperspective, but the customers
added.
But we need to respect those aswell when we suggest the
purchase line for this newlyreceived invoice.
(51:00):
So there are many tricky things, those as well.
When we suggest the purchaseline for this newly received
invoice, right.
So there are many, many trickythings.
Then comes we also don't knowthe industry of this company.
Okay, that might be easier tosolve.
We could just add a new fieldto company information.
You fill it in, you know, basedon some predefined list, maybe,
(51:23):
or maybe we just look up yourcompany name online little and
figure out what is the industryof your company, something like
that, if that's reliable enough.
Then comes things that are sortof user specific or role
specific.
We don't.
So obviously we know what therole the payable agents has, but
we don't really know whoeverturned on the payables agent,
(51:44):
what their role is.
That might be a person wearingall hats in the company that
turned it on.
It might be the accountant whoturned it on or some admin, ic,
admin person.
So there's so many things wherewe need to get a lot more
context, and those are thingsthat we're we're trying to work
on to figure out how can we makeBC more self-aware and
(52:05):
context-aware.
But this is just under the hood.
So many things.
I want to come back to one thingyou mentioned, brad, so I think
well, both of you mentioned it.
Actually, no doubt we see anagent as having a specific role
or specific business process tosort of fulfill, like handle
(52:30):
this venture, invoice processfrom start to finish.
Already, there we're able todefine some context for this
agent, to define some bounds forit to say, I mean, if it tries
to do something, we can sort ofrein it in or fence it in in a
way, because we know the context.
We don't imagine one agent, onesuper agent, that just does
(52:53):
everything, has all permissionsin the system, that can do
everything.
An agent needs to have someinherent knowledge about what is
it there for, and so we're,yeah, and that also has, of
course, some benefits in termsof permissions, because it's
it's, it's a user in the byitself.
You can set up the, the uh,what is what it has access to in
(53:16):
pc, and so on, um, but in terms.
So come back to your questionbrad about how to stitch all
these things together.
What is AL code and where doesthe agent instructions fit in?
And when do we ask the LLM andwhere do we feed it data If it
needs to search our purchasehistory?
(53:37):
Can we just loop through 1million deal entries or like all
these things?
And this is super hard.
The one thing that we'reexperimenting with now, because
we have our own agent frameworkinside BC.
That is not Copilot Studio.
We went our own way because wewere waiting for Copilot Studio
(54:00):
to bring some of thecapabilities that we need.
So right now, the sales orderagent and payables agent in BC
are built inside BC, if you will.
Those agents work in a waywhere they can navigate the UI.
So a bit like the client, likethe business central client that
(54:23):
you see in UI, imagine theagent being able to navigate
that UI.
So let's take a scenario whereyou receive an invoice.
You need to figure out just asuper simple example.
You need to figure out if thisbank account that is stated on
the invoice is known to us onthe vendor.
(54:44):
We could do it in AL.
We could just do a three linesof code in AL check loop through
the vendor bank accounts andcheck if it's a match.
Yeah, but that doesn't soundvery agentic.
In a way we could also tell theagent go to the vendor card, go
(55:05):
to the vendor bank accounts, gothrough the bank accounts one
by one and compare them to thisbank account you have on the PDF
.
If we can make that work asinstructions, like in a no-code
instructions, hey, go check ifthe vendor bank account is known
to us by going to the vendorcard, blah, blah.
Go check if the vendor bankaccount is known to us by going
to the vendor card, blah, blah,blah.
(55:25):
Then we've not written code Now.
We've written natural languagein an instructions prompt
somewhere.
So we always have thesedecisions to make and the agent
say do we just code it in AL ordo we try to push the agent to
do it via the UI?
And we try to do an agenticfirst motion here.
Because if we can succeed withthat then in time you can
(55:48):
imagine just going into BCpulling up a window, describe
the agent that you want and howit should behave in the UI and
then it would just go do allthese things, no AL code needed.
Basically, it would justnavigate the UI.
That is sort of the thinking ofour agents within BC.
Speaker 2 (56:07):
You said a key phrase in there, a key few-choice
words, where you said with theAL code, we're writing code for
a specific scenario and then nowwe can instruct the agent to
work with the UI.
Now we can instruct the agentto work with the UI.
So is the agent working withthe UI directly to perform its
(56:29):
actions, or is it working withcode behind the scenes?
I guess you could say in alow-code fashion of concrete
scenarios that it's goingthrough and trying to find, and
then also to the to tack ontothat, what about the variability
?
I'm all for agents havingspecific tasks and then they can
(56:52):
do those tasks.
Well, similar to what we'retalking we talked about at the
onset, where there's so manythings that we want to do or I'd
like to do, but I can't do itall because I don't have time.
Nor could I do be you know, forlack of better terms good at
anything if I do too much,because I'll have no time to be
able to focus on it.
So agents can have specifictasks.
How is it working?
(57:14):
And what about variablescenarios?
So today we talked about bankaccount.
Okay, now we can put it in, wecan teach it, and I do want to
go back to the UI and how itinteracts with it.
But what if tomorrow it's ourvendor number for a vendor?
Speaker 3 (57:31):
Yeah, let's maybe go to the UI piece first, because
just a quick bit of explanationof how it works and I might I'll
probably vastly oversimplifythings here and if some of our
engineers are listening to thisthey'll probably say, yeah, it's
a bit more detailed than that.
So think of a page object withall the metadata it has about
(57:58):
which fields are on it and so on.
Today, very simply put, allthat data is being exposed to
the client as some internals orweb service or APIs.
The client reads that andrenders the page right.
So that same API the agent nowhas access to.
(58:21):
So the agent can just notrender but read all that same
information that's on the page.
So when we say navigate to thevendor's page, it does so like
the client would.
In essence, there's no physicalclient spun up somewhere, it's
just the agent consuming thosesame APIs and navigating.
Seeing in code it's like thematrix.
(58:43):
You see all that green text.
The agent basically just looksat all the metadata and say, oh,
there's a caption that saysvendors there, let me try to
navigate to that.
So that's very simple.
But the way the agent can sortof navigate the UI, you know,
can say, hey, go to the linesand create a new line, add this
value here, and that's also whyit will always see what the user
(59:07):
would see, if it has the access, of course, to that.
So that also means that when webuild UI for the agent to
traverse, we need to removeambiguity and need to remove
stuff that makes it confusingfor the agent, which happens to
have the nice side effect thatwhat is confusing for agents is
(59:30):
probably also confusing forusers.
So if we can simplify some more, that's probably a good thing.
Yeah, so that's just about howthe agent will sort of work.
Will we sometimes have to dothings in AL instead?
Have to do things in AL instead, where, because it's just like,
(59:50):
for example, finding a certainrecord or things of that nature,
might make more sense in somecases in AL instead of having
the agent do it?
That's also a cost concern orconsideration we need to think
about.
Well, when the agent navigatesthe UI in this way, it uses the
(01:00:12):
LLM, so inherently we will havea cost of running the agent,
basically because that wewouldn't, in the same way, have
an AL, at least not.
Well, the machines where theNST is running is running anyway
, but still there are manyfactors in this, in deciding do
(01:00:32):
we go for AL, do we go fornatural language?
In essence, and we're trying topush more to go towards natural
language.
Sorry if I didn't really answeryour question there, but no you
did answer it and it's.
Speaker 2 (01:00:43):
I'm pleased to hear there's also a consideration
today, because no one knows whatthe future would be.
It's almost like I remember whenthe cell phone was invented and
it was probably $300 a minuteto use the cell phones.
Now cell phones are fairlyreasonably priced for unlimited
text and data for a given monthin the onset, while the
(01:01:14):
technology is still new and itcan be costly.
In essence, to see that thereare some considerations to
appropriately use the agent inconsideration for cost and
function, not just to use anagent for the sake of using an
agent.
So it is nice to hear thatthose considerations are in
there and I do understand nowthat it's working with the ui
and that's how we can limit whatit has access to and what it
(01:01:34):
can do within the context of theuser executing the agent and
what they have access to, if I'mif I understood correctly yes,
it's I've had conversation withsome and there may be a
misconception that you knowagents are just, you know,
co-pilot workflows or powerautomate flows or some fashion
(01:01:59):
of that, and it's nice to hearthe details, and I've been
having conversations witheveryone.
It's a lot more than that and isthat a true misconception?
Or is it just somebody tryingto knock AI similar to back?
When I see Chrissy, as we saythis almost every day, it's you
use the appropriate tool forwhat you're trying to do.
Asking AI to say five plus fiveand it and being able to make
(01:02:21):
it do 11 doesn't do anythingbecause you can use a calculator
, or maybe there should be anagent for math versus an agent
for payables.
So it's speak.
It's just like people or humans.
I speak to a specific human ifI need specific information.
You know an electrician, acarpenter or something.
I went on a little tangent there, but yes, you did answer my
(01:02:41):
question.
Speaker 3 (01:02:43):
And I think that this is worth just drilling a bit
more into like the role of anagent.
I know we always spoke aboutthat it has a certain role, uh,
to, to, to play like.
Uh, how that?
Like, if you draw up anorganization and you look at the
agent landscape like, what,what should an agent landscape
(01:03:04):
look like in in thisorganization?
Oh, do I need a payables agent?
Do I need a sales agent?
Do I need, like, when is myplate full of agents that now
all bases are covered in mycompany?
And how do they hand over toeach other?
And do I have a master agent ofall of them?
Or do I have, like that sort oforchestrates it all or are they
(01:03:25):
able to orchestrate betweenthem and just hand over?
And how small should they be?
Like, also, when you considerthat it's a user that they sort
of emulate in a sense, should weconsider that the classic sort
of segregation of duties, allthose things like do an agent
map to a human one-to-one orlike, and how does that look and
(01:03:49):
how does that look like in thisorganization?
So, and there are so manythings like, we've only just
begun, so I think that'ssomething we need to get wise on
as we go along.
Personally, I don't see onesuper agent that does everything
.
I would rather go smalleragents, much more limited scope,
because that also means we candefine their context much more
(01:04:09):
precisely.
Much more limited scope,because that also means we can
define their context much moreprecisely.
So they're.
They ideally wouldn't go out ofbounds of what they're supposed
to do so often.
But then you could imaginehaving like, let's say, you had
five finance agents.
Then you could have anoverarching supervisor agent
within finance that could helpsort of orchestrate between
these five finance agents.
(01:04:30):
Could also be you had a vendorcommunication agent that said oh
, now we get an invoice in, thisgoes to the payables agent.
But the next time there's anemail from the vendor it happens
to be a reminder that theyhaven't gotten our payment.
So obviously that doesn't go tothe payables agent, this goes
to someone else, another agent.
(01:04:51):
So who's orchestrating that ifthey're all monitoring that same
inbox?
That's stuff we need to figureout.
But the discussions havestarted.
How do we?
What does the agent landscapelook like?
How do they hand over?
Where do the humans fit in Allof that?
Where do we have the oneoverview to rule them all?
Speaker 1 (01:05:13):
and all that stuff.
You need that business uh,orchestrated or that
orchestrates all the agents.
Right, like if, if I'm abusiness owner and I need it to
do something, I I I'd ratherjust go to a single agent that
can help me orchestrate toaccomplish the goal of my
(01:05:35):
business, rather than, yeah, mygoal involves this and then I
got to go talk to individualagents, versus going to a single
maybe a tenant agent that livesin your tenant and it says I
need to do this so it may pullin certain points of my
(01:05:58):
applications, whether that'sbusiness central or maybe I need
to interact with teams andwhatever that is, it's much
easier for me from a businessperspective, much easier to go
to a single agent that willorchestrate that.
Speaker 3 (01:06:09):
And that's a great way of thinking about it.
And I think when I say that weI personally I'd like to
componentize agents as much aspossible, try them as small as
possible.
That doesn't mean you'll have100 interaction points as a user
.
You could still have that oneinteraction point, as you say,
like the master orchestratoragent, or maybe within that
domain you work as a user, andthen that agent will, will,
(01:06:30):
coordinate with all the otherusers.
So it will, it will havedirects, if you will like.
Uh, so it's.
It's a bit the same kind oforganization, yeah, same
hierarchy, if you will.
And so, and I think that that's, that's just some
technicalities.
We need to figure out how doyou interact with an agent.
It doesn't necessarily need tobe a one-to-one representation
(01:06:52):
of how are the agentsrepresented in the system under
the hood somehow.
So, uh, you could have a one,one to twenty relationship, uh,
between you and agents, butthere could be a single pane of
glass that will be thatinterface or whatever that would
be super fascinating.
Speaker 1 (01:07:11):
It would almost really like it's a one-to-one
human, like a true co-pilot,right.
Like I'll have a co-pilot of aVP of operations, right.
Like that would be the personthat I would go to and say I
need to accomplish this, how doI accomplish this in the system.
And then that agent for thatspecific role for operations
(01:07:32):
would then just interact withall the different agents.
That also kind of follows thehierarchy within your
organization.
That's how maybe I'd see it,because then every role has a
co-pilot.
Speaker 3 (01:07:47):
No, to me that makes perfect sense.
And I think where we are now isthat we need to start with we
need to also be mindful of wherewe are and not try to build
100% automation to begin with.
We need to learn and there'sbenefits in involving users more
(01:08:11):
often to begin with, becausethere's a lot of learning
moments when users are involvedWhenever, as I mentioned 30
minutes ago, whenever the userchanges some of our suggestions
or the agent's suggestions,that's a learning opportunity.
So there's sort of great valuein having human in the loop and
I think if we can combine itwith the risk of being a little
(01:08:34):
bit philosophical but if we canmake humans and agents work
together and complement eachother's weak sides, that's that
we would have a verywell-working agent.
Aiming for 100, I think.
Think, of course I mean wewouldn't want accounts payables
to be 100% if it's reliable andfast and well, maybe speed
(01:08:58):
doesn't matter as much, but atleast reliable.
But I think that will take timeuntil we get there and that's
fine.
If we can solve for the 75% ofthe normal day-to-day use cases,
that's also super cool.
And then humans can do the morecomplex handle, the more
(01:09:20):
complex invoices and over time,as we get better using this
technology, the agents can getbetter.
Then probably we can increasethat to I don't know 80, 90
percent, who knows if we getclose to 100.
Speaker 2 (01:09:37):
There's so much in there and I think it's, you know
, the future.
Who knows where the future willbe?
And again, I do think we're atthe point just in our society,
just to go philosophical for alittle bit because,
realistically, some of this AIagent stuff, with it being so
new and learning what it can doand seeing how it reacts,
responds and works, isphilosophical because we don't
(01:10:01):
really know.
But it would just beinteresting to see where it will
all be and how this will allwork uh, in the years.
And I think now we're justthrowing a lot of agents against
the wall to see which work,which stick, which are effective
, and then also how to, to yourpoint, maybe even break them
down to have separate agentsthat have specialties or
(01:10:25):
specializations, to be able tohave them work together.
Speaker 1 (01:10:29):
You know, I think it's going to be.
I would love to see it more of arole specific, because, if you
look at the way I use agents orco-pilot, right, for example, in
my day-to-day it's typicallyaround my role, and so it would
be interesting to see if this ismaybe where you guys are
(01:10:52):
heading toward, where Microsoftis going towards, where there is
a co-pilot working along withme, because then I can have my
conversation with my CEO or yourpresident or anybody else in
your organization is morestrategic, strategic, right.
There's something that us humansare more creative and more
(01:11:13):
strategic about things, that weyou know what we do in the
business, where the hard, youknow, the tedious work of
getting to do things is wherethe co-pilot comes in, where,
hey, I need some data, you know,give this information to me
while we're trying to make adecision of where the business
is going.
I think that would make a lotmore sense that you'd always
have a co-pilot with you, basedon the role and, of course,
(01:11:35):
structurally in terms ofhierarchy within the system.
It also makes sense because weunderstand that your agents are
based upon a role, right,because it's UI driven.
I think it's limited to aspecific role, and so maybe
that's where it is, because nowyou have, you know, an AP
payables agent, now you didaccounts receivable agent and so
(01:11:56):
forth, sales agent and so forth.
Maybe that is where thestructure is.
To me it makes more sensebecause my area of focus is just
that's, my interaction anyway.
Speaker 3 (01:12:09):
No, and I think that makes perfect sense, and I think
that this is why we've allchosen this path and chosen sort
of well-known scenarios, roleseven you can say they're even
more narrow than a role at thispoint, because a payables agent,
like a real accountant, mightdo more, or an AP clerk might do
(01:12:29):
more than just registerinvoices.
I guess it depends also percompany, and I think what we're
doing now is a lot ofexperimentation, and I think
that's the only way forward.
As you said, brad, like we tryto create some agency which ones
work as the technology improves, as it does every day, and I
think we also just need to behonest that I mean, as I said,
(01:12:52):
two and a half years ago thisfield didn't exist, so there are
, by definition, no experts.
We're also learning, right,we're just working on this all
day, every day, so we get a lotof exposure to it as
technologists.
So we get a lot of exposure toit as technologists, but experts
(01:13:13):
, I mean I would like to saywe're getting there, but you
know who can be an expert in twoand a half years of exposure to
something?
So this will take time, sowe're all learning right and an
expert on something that'schanging daily.
Speaker 2 (01:13:28):
So it's not only that it's new.
Generally speaking, it's in itsinfancy.
But it seems every time I wakeup and put on the morning
finance news, they're talkingabout some new LLM, local large
language model, some new agent,some new version that it's very
difficult for anyone to keep up,which is also important to
(01:13:51):
mention.
Some are talking about that itmay be too late to learn or too
late to get into it, buteverybody's in the same boat
because everything's new.
So no one has missed the agentexperimentation phase, even
Copilot Studio, as you hadmentioned.
Speaker 3 (01:14:10):
they keep adding features and changes, so and get
in and start working with ityeah, and this is, of course, a
much larger challenge that we,like the industry, face, and all
companies out there wants toconsume all companies out there
who wants to consume AI or toconsume features within their
(01:14:32):
ERP system or whatever.
So, on one side, we have ahigher than ever before desire
to experiment, test out things AB test, try out things with
real customers and so on, andthat will only increase Our
(01:14:53):
desire to do that will onlyincrease, and we need that
because technology is changingEvery day.
We look in the toolbox, toolslook different, work different.
So on the other side of that,you have customers who need
predictability, they don't wantto be disrupted, they value
business continuity, all thesethings.
(01:15:13):
So how do these two ends meet?
We need experimentation andwhat you call fail fast or adapt
fast or pivot fast Customers.
They just wanted to work thesame way as they did yesterday.
So how do they meet?
And I think that comes back to,at least for Business Central
(01:15:33):
specifically, we need to figureout ways where we can create
test rings and let people optinto being early preview testers
or maybe test experimentalfeatures in a sandbox or what
have you, so we can get somesome feedback.
Yeah, but this is, this is achallenge, for sure.
(01:15:54):
And also when it comes toadoption, as we spoke about an
hour ago, like if it doesn'twork or not in terms of a bug,
but if I don't get the resultfor of a feature that I expected
of this agent, will I abandonit or will I give it a new shot
tomorrow where we've maybechanged some stuff and updated
(01:16:16):
it and all of that.
You had a great episoderecently about change management
, which I loved.
That was just.
All of these things are justmore important than ever, and I
think we have some part of thatthat we need to do Educate,
explain value even in the UI,make it evident what is the
(01:16:40):
value for you as a person, notonly for the company, but also
for the company.
This is going to be even moreimportant than ever and super
hard.
Speaker 1 (01:16:51):
Yeah, I think I mean just thinking about just going
back to the hierarchy componentand you brought up the adoption.
You know, a lot of times youknow employees or users of an
application or even within thebusiness itself.
It typically bleeds from whatyou do on a personal side of
things and then you kind ofbring that into your business.
(01:17:11):
So, for example, I'd call that.
You know I use Copilot orChatGPT or whatnot and when you
chat right, you have differentchats.
Typically, when you're creatingyour chat you ground it of what
that role is.
You are my finance manager andI have this situation and I need
(01:17:36):
you.
It's almost like that, right,with Business Central.
I would think that's wherewe're heading.
I can't foresee that.
I'm just thinking out loud herewhere I would love to structure
it that way where what I do nowpersonally I could do the same
thing in Business Central.
I'm going to add an agent.
You are this agent.
I need you to do this.
(01:17:57):
It's almost in that way of easyfor people to kind of
understand because that's whatthey do.
Now I mean, it's becoming aday-to-day occurrence.
I chat with my chat GPT or grokor whatever AI that you use for
chat.
But again, I can't predict that.
Just the way I do things in aday-to-day, it may bleed into
(01:18:21):
Business Central.
I don't know.
Speaker 3 (01:18:24):
Yeah, but I think it makes sense.
And if you think about the waythat the agent, like you, have a
bunch of agents enabled foryour organization.
They work for different teams,for different users and so on,
and at the end of the day, therewould be supervisors of agents,
like human supervisors ofagents, that would work with
(01:18:47):
agents, make sure that thewheel's always spinning, that
the agents are not stuck onanything, of course, with all
the business approvals built inthat you typically want.
When you need to approve a newvendor that's great in the
system, or need to send the POfor approval, all that stuff
still works as it used to.
(01:19:07):
The agents don't change that.
All that stuff still works asit used to right the agents
don't change that.
It might be that you can haveagents that will help these
approvers as well.
Like you can imagine an agentthat says you know, here's a new
approval request from Chris whowants to approve this purchase
order of $10,000.
Well, based on history, there'snothing to worry about.
(01:19:30):
You can easily go approve thisone.
Should I go approve it the nexttime on your behalf?
There are many applicationswhere you can think about agents
kicking in and learning fromthe history and, looking at you
know, try to identify risk,because if there's no risk, why
stop at all?
Why not just go ahead and andum, yeah, but let's see how far
(01:19:53):
we can.
We can drive this automationbecause it is automation brad
you asked about.
Was this just a glorified flow?
Depends on how you look at it.
I mean, from a conceptualstandpoint, you could say yeah,
uh, but it's a flow that's ableto handle much more variability,
in a sense, like where flow isvery deterministic.
(01:20:15):
You put a stick in the wheel ona certain point it just breaks
and falls apart.
Where here it's a bit so yeah.
A guy said to me some months agowell, this is like when you
build houses in areas where youhave a lot of earthquakes you
(01:20:35):
build them so the foundation cansort of move and the house can
move.
It's a bit like the same thing.
You need to build in somewiggle room so the agent can
take some decisions within acertain context, within certain
constraints and so on, but notbreak when something doesn't
follow the sunshine pathUnderstood.
(01:20:57):
It's not a bunch of ifs andthens yes, it's.
Speaker 2 (01:21:02):
No, it's.
I mean, I guess, ultimately, Iguess any workflow is a bunch of
ifs and thens.
It's just a matter of which ifsand thens and whens you can
react to, or respond to withoutbreaking.
So we've talked a lot about thepayables agent.
What is the current state ofthe payables agent as far as
(01:21:24):
availability and what are theshort-term future plans?
I know nobody can plan too farinto the future because
something could change tomorrow,but to anybody listening, we've
been talking about it.
What is the current state andavailability for it?
Speaker 3 (01:21:39):
Yeah, great question.
So because the agent navigatesthe UI, as we've talked about,
we've had to improve some stuffin sort of the base application,
especially around e-documents.
We've had to build a connectorwhere basically JobQ, or the
agent, via the JobQ, could pullin an invoice that comes in on a
(01:22:03):
certain mailbox or on aSharePoint folder and so on.
So we've done some componentsthat the agent will use and
those are right now, and we'llcall sort of private preview
where if you're on a certainlist, your tenant is on a list,
then you can get access to someof these components and try them
out.
This is not in any way anyagentic experience yet.
(01:22:26):
This is just some improvementsto the base product and we'll
keep doing those over the nextsort of few months and the idea
is that we'll add the first sortof version of the agentic
experience, enabling theend-to-end flow, creating the
invoice and so on, with the nextmajor release.
That is the next.
(01:22:47):
That is the plan.
So end of year, this calendaryear, basically, how much will
it then support in terms of howmuch accounting knowledge we'll
have?
Well, that's what we need tofigure out from now and until
then.
Will we have the policy conceptworking?
(01:23:08):
Will we have some learningconcepts working Well?
How much would we look atpurchase history when we try to
determine what the invoiceshould look like, and so on?
So that's to be decided.
But ideally we would have thefirst version working with the
next major release and in US andmaybe some other
(01:23:31):
English-speaking countries tobegin with, and then it's going
to be an iterative process,adding more capabilities, adding
more countries and so on afterthat.
But yeah, keep an eye out forthe next major release.
Speaker 2 (01:23:48):
That's the current plan Always a lot of great
things added to each of themajor releases, and well, so we
appreciate you taking the timeto speak with us today.
It's always a pleasure speakingwith you, whether it's either
on the podcast or in person, oreven in some remote
conversations.
We do appreciate you takingyour time to speak with us and
we appreciate all that you dofor the community and as well as
(01:24:10):
all of what your team is doingfor Business Central for the
users of the application.
If anyone has any questions orwould like to learn more about
the Payables, agents or otherinitiatives that you're working
on, what's the best way to getin contact with you?
Speaker 3 (01:24:22):
Well, this moment there's probably two good ways.
So one is on the Yammer site orViva Engage If you're a partner
.
Reach out there If you havequestions that you don't want
the world to see.
Otherwise, feel free to hook upwith me on LinkedIn.
It's the only place that I'msort of active from a sort of
social media perspective, sofeel free to send me messages
(01:24:44):
there or tag me in the post orwhatever, and I'll be there.
Speaker 2 (01:24:46):
Great, Thank you very much for taking the time to
speak with us again, and I hopeyou get some great gardening
into the season as well.
Speaker 3 (01:24:55):
Thank you so much, guys, pleasure all right, thank
you so much soon ciao, ciao, bye.
Speaker 2 (01:25:00):
Thank you, chris, for your time for another episode
of in the dynamics, corner chair, and thank you to our guests
for participating thank you,brad, for your.
Speaker 1 (01:25:09):
It is a wonderful episode of Dynamics Corner Chair
.
I would also like to thank ourguests for joining us.
Thank you for all of ourlisteners tuning in as well.
You can find Brad atdeveloperlifecom, that is
D-V-L-P-R-L-I-F-E dot com, andyou can interact with them via
(01:25:29):
Twitter D-V-L-P-R-L-I-F-E dotcom, and you can interact with
them via TwitterD-V-L-P-R-L-I-F-E.
You can also find me atMattalinoio, m-a-t-a-l-i-n-o dot
I-O, and my Twitter handle isMattalino16.
And you can see those linksdown below in the show notes.
(01:25:50):
Again, thank you everyone.
Thank you and take care.
Popular Podcasts
On Purpose with Jay Shetty
I’m Jay Shetty host of On Purpose the worlds #1 Mental Health podcast and I’m so grateful you found us. I started this podcast 5 years ago to invite you into conversations and workshops that are designed to help make you happier, healthier and more healed. I believe that when you (yes you) feel seen, heard and understood you’re able to deal with relationship struggles, work challenges and life’s ups and downs with more ease and grace. I interview experts, celebrities, thought leaders and athletes so that we can grow our mindset, build better habits and uncover a side of them we’ve never seen before. New episodes every Monday and Friday. Your support means the world to me and I don’t take it for granted — click the follow button and leave a review to help us spread the love with On Purpose. I can’t wait for you to listen to your first or 500th episode!
Crime Junkie
Does hearing about a true crime case always leave you scouring the internet for the truth behind the story? Dive into your next mystery with Crime Junkie. Every Monday, join your host Ashley Flowers as she unravels all the details of infamous and underreported true crime cases with her best friend Brit Prawat. From cold cases to missing persons and heroes in our community who seek justice, Crime Junkie is your destination for theories and stories you won’t hear anywhere else. Whether you're a seasoned true crime enthusiast or new to the genre, you'll find yourself on the edge of your seat awaiting a new episode every Monday. If you can never get enough true crime... Congratulations, you’ve found your people. Follow to join a community of Crime Junkies! Crime Junkie is presented by audiochuck Media Company.
Ridiculous History
History is beautiful, brutal and, often, ridiculous. Join Ben Bowlin and Noel Brown as they dive into some of the weirdest stories from across the span of human civilization in Ridiculous History, a podcast by iHeartRadio.